Extended Abstracts Book PDF

Alexis Quesada-Arencibia
Roberto Moreno-Díaz jr.

José-Carlos Rodríguez (Eds.)
EURO CAST 2005
Cast and Tools for

Robotics,
Vehicular and
Communication
Systems
EXTENDED ABSTRACTS
10th Internacional Workshop on Computer Aided Systems Theory

Las Palmas de Gran Canaria, Spain, February 2005
Tenth International Conference on
COMPUTER AIDED SYSTEMS THEORY
EUROCAST 2005
Edited by
Alexis Quesada-Arencibia
Roberto Moreno-Díaz jr. and
José Carlos Rodríguez-Rodríguez
© 2005 IUCTC Universidad de Las Palmas de Gran Canaria
ISBN: 84-689-0432-5
REGISTRO: 05/1325
Preface
The concept of CAST as Computer Aided Systems Theory, was introduced by F.
Pichler in the late 80's to encompass those computer theoretical and practical
developments as tools to attack problems in System Science. It was thought as the
third component (the other two being CAD and CAM) that will provide for a
complete picture of the path from Computer and Systems Sciences to practical
developments in Science and Engineering.
The University of Linz organized the first CAST workshop in April 1988, which
demonstrated the acceptance of the concepts by the scientific and technical
community. Next, the University of Las Palmas de Gran Canaria joined the
University of Linz to organize the first international meeting on CAST, (Las Palmas
February 1989), under the name EUROCAST'89 and proved to be a very successful
gathering of systems theorists, computer scientists and engineers from most of
European countries, North America and Japan.
It was agreed that EUROCAST international conferences would be organized every

two years, alternating between Las Palmas de Gran Canaria and a continental Europe
location. Thus, successive EUROCAST meetings took place in Krems (1991), Las
Palmas (1993), Innsbruck (1995), Las Palmas (1997), Vienna (1999), Las Palmas
(2001) and Las Palmas (2003) in addition to an extra-European CAST Conference in
Ottawa in 1994. Selected papers from those meetings were published by Springer-
Verlag Lecture Notes in Computer Science nos. 410, 585. 763,1030, 1333, 1798,
2178 and 2809, and in several special issues of Cybernetics and Systems: an
lnternational Journal. EUROCAST and CAST meetings are definitely consolidated, as
it is shown by the number and quality of the contributions over the years.
EUROCAST 2005 continue with the approach tested last Conferences (2001 and
2003). There are different specialized Sections which, in this occasion, are devoted to
Formal Approaches in Modelling (C1), Tools and Models for Social Systems (C2),
Intelligent Information Systems (C3), Systems Applications Components (C4),
Information Applications Components (C5), Cryptography and Spectral Analysis
(C6), Computer Vision (C7), Biocomputing (C8), Intelligent Vehicular Systems (C9),
Robotic Soccer (C10) and Robotics and Control (C11).
The chairpersons, with the counselling of the International Advisory Committee,

selected the 106 extended abstracts for oral presentation at the meeting which are
included in this volume. Specific instructions will be given for the preparation of the
Post-Conference Publications, which contain full papers selected, as in prior
EUROCAST's, after the oral presentations. The volume is divided in the eleven
Sections explained above.
The event has been possible thanks to the efforts of the chairmen of the Workshops in
the selection and organization of all the material. The organizers must express their
acknowledgement to all contributors and participants, to the invited speakers for their
readiness to collaborate, to the sponsors and in sum, to all those which will contribute
to the success of the Conference as a pleasant and useful scientific, technical and
human event.
Table of Contents
I. Formal Approaches in Modelling
On the formal and physical frontier areas between human

knowing and machine knowing.....................................................................1
J. Mira
Approximation Problems Categories.............................................................3

L. Aparecida-dos-Santos, D. Moraes-Claudio, L. Vieira-Toscani,
P. Blauth-Menezes
Computation of Partial Automata through Span Composition ......................5

K. Girardi-Roggia, M. Augusto-Hoff, P. Blauth-Menezes
Degenerate Arrays.........................................................................................7
M. Miró-Julià
Neural network sensitivity analysis applied for the estimation of

sensor matrix redundancy............................................................................11
P. Szecówka, A. Szczurek, B. Licznerski, F. Pichler
Fuzzy modeling for coal seams A case study for a hard-coal

mine.............................................................................................................13
L. Argüelles, J. M. Rivas, J. Toraño, C. González, R. García, T.
De-Pedro, J. A. Martín
Data Analysis of a Distributed Solar Collector Field ..................................15

M. Berenguel, R. Klempous, H. Maciejewski, J. Nikodem, M.
Nikodem, L. Valenzuela
Non-linear Formal Modelling of Amphibian Tectal Cells ..........................17

G. de Blasio, R. Moreno-Díaz
Optimization of a Class of Uncertain Systems Based on

Uncertain Variables.....................................................................................21
Z. Bubnicki
Computational Simulation of Categorical Constructions ............................23

R. Born-Vieira, P. Blauth-Menezes
Composing Transitions Into Transactions in UML Diagrams.....................25

J. Pereira Machado, P. Blauth-Menezes
i
One formal systemic approach in software design. Case study:
Extreme programming.................................................................................29
C. Halbich
On the Application of Biological Vision Concepts and

Mechanisms in very Demanding Industrial Visual Processing ...................33
A. Quesada-Arencibia, J. C. Rodríguez-Rodríguez, R. Moreno-
Díaz jr
II. Tools and Models for Social Systems
Theory-Building with System Dynamics: Principles and

Practices ......................................................................................................36
M. Schwaninger, T.K. Hamann
Ontology Integration for Statistical Information .........................................41

W. Grossmann, M. Moschner
Solving Influence Diagrams With Simulation Methods..............................43

A. Moreno-Díaz, M. Virto, J. Martín, D. Ríos-Insua
Decision support system using data warehousing and on-line

analytical processing technologies ..............................................................45
R. Klempous, H. Maciejewski, J. Nikodem, M. Nikodem, J.
Rozenblit, J. Scharinger, C.P. Suarez Araujo
Towards Automated Controlling of Human Projectworking

Based on Multiagent Systems .....................................................................47
M. Mauerkirchner, G. Hoefer
III. Intelligent Information Systems
On Recursive Functions and Well–founded Relations in the

Calculus of Constructions............................................................................49
J. L. Freire, E. Freire, A. Blanco
Longest Sorted Sequence Algorithm for Parallel Text Alignment..............51

T. Ildefonso, G. Pereira
Information Retrieval and Large Text Structured Corpora .........................55

F. M. Barcala, M. Molinero, E. Domínguez
ii
Meteorological Image Descriptors ..............................................................58
J. L. Crespo, P. Bernardos, M.E. Zorrilla, E. Mora
Matching with a Hierarchical Ontology ......................................................60

Y. Choueka, N. Dershowitz, L. Tal
Business Intelligence Process through Information Extraction

with Lixto ....................................................................................................64
J.A.. Díaz-Prado
Reasoning about a Functional Program for computing Gröbner

Basis ............................................................................................................69
J. Santiago-Jorge, V. Gulías, J. L. Freire, J. J. Sánchez
CheapTB: A Low Cost of Operation Distributed File System ....................71

J. París, V. Gulías, C. Abalde
Spelling Correction on Technical Documents.............................................73

M. Vilares, J. Otero, J. Graña
Verification of Language Based Fault-Tolerance........................................77

C. Benac-Earle, L. Fredlund
Applying Stacking and Corpus Transformation to a Chunking

Task .............................................................................................................80
J. A. Troyano, V.J. Díaz, F. Enríquez, V. Carrillo
A Formally Verified Proof (in PVS) of tje Strong Completeness

Theorem of Propositional SLD-resolution ..................................................83
J.A. Alonso, M.J. Hidalgo, F.J. Martín-Mateos, J.L. Ruiz-Reina
Fuzzy ontologies generation in GUMSe .....................................................87

J. Serrano-Guerrero, J.A. Olivas, J. De-la-Mata
Extracting Computer Algebra programs from statements ...........................89

J. Aransay, C. Ballarin, J. Rubio
Cross-Language Classification of Function Words .....................................92

P. Gamallo, G.P. Lopes, J.F. Da Silva
Integrating syntactic information by means of data fusion

techniques....................................................................................................96
F.J. Ribadas, J. Vilares, M.A. Alonso
iii
An Open Low-Cost Parallel Architecture for Unsupervised
Learning in Information Retrieval.............................................................100
R Mones, N Vázquez, I. Díaz, J. Ranilla, E.F. Combarro, E.
Montañés
An Iterative Method for Mining Frequent Temporal Patterns...................102

F. Guil, A. Bosch, A. Bailón, R. Marín
IV. Systems Applications Components
Data Mining with Scatter Search...............................................................106

I.J. García-del-Amo, F. García-Lopez, M. García-Torres, B.
Melián-Batista, J.A. Moreno-Pérez, J.M. Moreno-Vega, R.
Rivero-Martin
Web Usage Mining Project for Improving Web-based Learning

Sites ...........................................................................................................110
M.E. Zorrilla, E. Menasalvas, D. Marín, E. Mora, J. Segovia
Similarity Queries in Data Bases Using Metric Distances - From

Modeling Semantics to its Maintenance....................................................113
J. Küng, R. Wagner
A WEB-CASE Tool Prototype For Hybrid Software

Development .............................................................................................115
F.J. Orellana, F. Guil, I.M. Del Aguila, S. Túnez
An Augmentative Communication System based on Adaptive

Evolutionary Hypermedia Systems ...........................................................119
M.V. Hurtado, N. Medina, L. García-Cabrera, M.L. Rodríguez
A tool for thematic cartography of corpora ..............................................123

T. Roy, P. Beust
System-oriented Evaluation for Multi-tier Personalization of

Web Contents ............................................................................................127
A. Díaz, P. Gervás, A. García
The gaps of the thesaurus WordNet used in Information

Retrieval ....................................................................................................131
J. De-la-Mata, J.A. Olivas, J. Serrano-Guerrero
iv
V. Information Applications Components
Language for Uncertainty and Conflicting Agents....................................134

G. Resconi, J. Alonso, R. Izquierdo
A Model-Based Architecture for Fuzzy Temporal Diagnosis ...................136

J.M Juárez, J. Palma, M. Campos, J Salort, A. Morales, R.
Marín
Ontology-Based Models of Human-Computer Interaction for

Usability Evaluation and Usable Design ...................................................139
E. García-Barriocanal, M.A. Sicilia
Extension of Ontologies assisted by Automated Reasoning

Systems .....................................................................................................143
J. Borrego-Díaz, A.M. Chávez-González
Applying Genetic Algorithms to Referring Expression

Generation and Aggregation......................................................................145
R. Hervás, P. Gervás
A Software Architecture for Effective Document Identifier

Reassignment ............................................................................................150
R. Blanco, A. Barreiro
An Ontology for Reusing Synthetic Tasks ................................................152

A. Rodríguez-Rodríguez, F. Quintana-Domínguez
Modelling and Simulation of Systems with Multi-level

Anticipation and Retardation.....................................................................155
D.M. Dubois, S. Holmberg
A Tractable Subclass of Fuzzy Constraint Networks ................................157

A. Bosch, F. Guil, R. Marín
Parallel State Space Generation and Exploration on Shared-

Memory Architectures...............................................................................161
M. Ceska, B. Krena, T. Vojnar
VI. Cryptography and Spectral Analysis
Tree-structured Legendre multi-wavelets..................................................165

E. Pogossova, K. Egiazarian, A. Gotchev, J. Astola
v
Remarks on Calculation of Autocorrelation on Finite Dyadic
Groups by Local Transformations of Decision Diagrams .........................169
R. Stankovic, M. Karpovsky
A new Pseudo-Random Generator based on Gollmann Cascades

of Baker-Register-Machines......................................................................173
D. Jochinger, F. Pichler
An Excellent Permutation Operator for Cryptographic

Applications ..............................................................................................176
J. Scharinger
Fault Cryptanalysis of ElGamal and Related Signature Schemes .............178

J. Biernat, M. Nikodem
Complexity-Theoretical Approaches to the Design and Analysis

of Cryptographical Boolean Functions......................................................180
J. D. González-Cobas, J. A. López-Brugos
Algorithm for Proving the Knowledge of an Independent Vertex

Set..............................................................................................................182
P. Caballero-Gil, C. Hernández-Goya
Improvement of the Edit Distance Attack to Clock-Controlled

LFSR-Based Stream Ciphers.....................................................................184
P. Caballero-Gil, A. Fúster-Sabater
Protocol Analysis for Concrete Environments ..........................................186

D. Gollmann
VII. Computer Vision
Pattern Recognition in AVHRR Images by means of Hibryd and

Neuro-Fuzzy Systems................................................................................189
J.A. Piedra, F. Guindos, A. Molina, M. Canton
From Projective Virtual Reality to the Virtual Human: Virtual

Worlds become realistic ............................................................................192
J. Rossmann
Multimedia Infrastructure for Intelligent Diagnosis and

Processing of Phonocoradiographic Signals..............................................194
J. Martínez, J. Palma, M. Breis, K. Herzog, A. Fritz, J. Lopez
Dactylology Language Translator from Video to Text .............................197

A. Camino, C.M. Travieso, J.B. Alonso, M.A. Ferrer
vi
Automatic Braille Character Recognition: BrailLector .............................201
N. Falcón, C.M. Travieso, J.B. Alonso, M.A. Ferrer
Retinal Based Authentication via DistributedWeb Application ................205

C. Mariño, M.G Penedo, M Penas
Medial Axis Transform Computation Using Shock Wavefront

Detection ...................................................................................................209
R. Cárdenes, J Ruiz-Alzola
Comparative analysis of Calibration Methods for a Static

Camera ......................................................................................................212
J. Isern-González, J.D. Hernández-Sosa, A.C. Domínguez-Brito,
A.M. Naranjo-Cabrera
Cue Combination for Robust Real-Time Multiple Face Detection

at Different Resolutions.............................................................................215
M. Castrillón-Santana, O. Déniz-Suárez, C. Guerra-Artal, J.
Isern-González
Evolutionary Color Constancy Algorithm based on the Gamut

Mapping Paradigm ....................................................................................219
C. Munteanu, A. Rosa, M. Galan, E. Rubio Royo
Vision based Automatic Occupant Classification and Pose

Recognition for Smart Airbag Deployment...............................................223
Min-Soo Jang, Yong-Guk Kim, Sang-Jun Kim, Jeong-Eom Lee,
Seok-Joo Lee, Gwi-Tae Park
VIII. Biocomputing
Wiener-type neuronal model with refractoriness .....................................227

V. Giorno, A.G. Nobile, L.M. Ricciardi
On Myosin II dynamics: from a pulsating ratchet to a washboard

potential.....................................................................................................229
A. Buonocore, L. Caputo, E. Pirozzi, L.M. Ricciardi
Feedback effects in Stein's coupled neurons .............................................231

A. Di Crescenzo, B. Martinucci, E. Pirozzi
Upcrossing first passage times for correlated Gaussian processes ............233

V. Giorno, A.G. Nobile, E. Pirozzi
vii
Convergence of Iterations..........................................................................235
P. Cull
Semiautomatic Snake-Based Segmentation of Solid Breast

Nodules on Ultrasonography.....................................................................238
M. Alemán-Flores, P. Alemán-Flores, L. Álvarez-Leon, M. B
Esteban-Sánchez, R. Fuentes-Pavón, J. M. Santana-
Montesdeoca
Parallel Progressive Multiple Sequence Alignment ..................................242

E. Pitzer
Computer Aided Simulation of Energy Limits and Development

of Bio Thermodynamic Systems ...............................................................245
S. Sieniutycz
IX. Intelligent Vehicular Sytems
Soft Computing and Geometrical Control for Computer Aided

Driving ......................................................................................................249
J. Alonso-Ruiz, T. De-Pedro, C. González, R. García
A Monocular Solution to Vision-based ACC in Road Vehicles................251

J. Nuevo, M. A. Sotelo, M. Ocaña, L.M. Bergasa
Automatic Parallel Parking for Autonomous Vehicles..............................255

J. De Lope, D. Maravall
Intelligent Automation of Electric Power Steering Systems .....................257

J.E. Naranjo, C. González, R. García, T. De-Pedro
Development of an Artificial-Vision Controlled Traffic-Light

Prototype ...................................................................................................259
L.J. Rodríguez-Aragón, A. Serrano, C. Conde, R. Montes-Diez,
E. Cabello
Permanency Effect in Scene Depth Analysis ............................................261

M.A. Fernández, J.M. López-Valles, A. Fernández-Caballero,
M.T. López, J. Mira, A.E. Delgado
Pedestrian Detection for Intelligent Vehicles based on Active

Contour Models and Stereo Vision ...........................................................265
C. Hilario, J.M. Collado, J.M. Armingol, A. De-la-Escalera
viii
Fast Road Sign Detection using Hough Transform for Assisted
Driving of Road Vehicles..........................................................................269
M. A. García, M. A. Sotelo, E. Martín-Gorostiza
X. Robotic Soccer
Advances in Robotics................................................................................272
P. Kopacek
Current and Future Trends and Challenges in Robot Soccer.....................275

N. Weiss, B. Reusch, L. Hildebrand
Strategy and communication in robotic soccer game ................................277

B. Horak, M. Obitko, J. Smid, V. Snasel
Rete algorithm applied to robotic soccer ...................................................280

M. Palomo, F.J. Martín-Mateos, J.A. Alonso
Vision system design and accompanying problems in robot

soccer.........................................................................................................284
G. Klancar, D. Matko, R. Karba
Time-optimal collision avoidance of automatically guided

vehicles......................................................................................................286
A. L. Schoute
XI. Robotics and Control
Towards a Biomathematical Model of Intentional Autonomous

Multiagent Systems ...................................................................................290
B. Mitterauer, J. Pfalzgraf
On a Hybrid Symbolic, Connectionist Approach for Modeling

Robot Arm Kinematics..............................................................................292
J. Pfalzgraf
Controller Network for a Humanoid Robot...............................................294

E. Schierer, M. Wuerzl
Programming by Integration in Robotics ..................................................297

J. L. Fernández-Pérez, A.C. Domínguez-Brito, J.D. Hernández-
Sosa, J. Cabrera-Gámez
ix
A Homeostatic-Adaptive Approach for Controlling Robotic
Systems .....................................................................................................301
J.D. Hernández-Sosa, J. Lorenzo, A.C. Domínguez-Brito, M.
Castrillón-Santana
A Mathematical Formalism for the Evaluation of C-space for

Redundant Robots .....................................................................................305
R. Therón, V. Moreno, B. Curto, F.J. Blanco
Global Modal Logics for Multiagent Systems: A Logical

Fibering Approach.....................................................................................309
J. Edtmayr
The H2 Control Problem:Comparison of State-space and

Transfer-function Approaches...................................................................311
V. Kucera
Analysis, Synthesis and Mimetization of Nonlinear Dynamical

Control Systems using Mathematica .........................................................316
J. Rodríguez-Millán, C. González, A. Patete
Embedded Control Systems: the Control Kernel.......................................321

P. Albertos, M. Vallés, A. Crespo
Contribution to the Control of a Distributed Solar Collector Field ...........325

M. Berenguel, R. Klempous, H. Maciejewski, M. Nikodem, J.
Nikodem, I. Rudas, L. Valenzuela
Author Index............................................................................................327
x

!

"
#
$

%
$

&%
' "&'#

%

( "# ) " # %

$ "!#
"*#

+
,
-./& "0# 1

" #

2 !3

"
# 2*03
24563 " 7 89
:89# '
%

)

;

&

" ! #
' ( "#
" < <#
$ "

# " #

"

%

#
=
&'

(
&' <

<
%

%

! >

*
"# %

<<

!

" # $% & ' !
(

) % *
+,,-.
/
# 012
34
56
7 2

*6
8

9
: 0
;2 82 :
5 +. 56
* 82 < =(=> 92*
(?

(? +/@@=.
=
# '&
4
2 1 8
4
2 (
6) /@@> 9
+/@@>.< A(=
> 4 '2 34 B: ) %
+,C.< (/@
- 'B ?

)
%

+,C/.
D E5 '2 F
%
2 B
:
< 2* 2 B
:
F
(
%
) ( 2

A +,A-.< ==(==/
A ? & '9

!

*) 2 2 E

"(
* 5 2 2 E 4 GH +,A,.
C
# ' 8 2 2
?
4

2 *6
2 *) 9 2 :2
8
8*6
* : / +,CA.< C,(C,-
Approximation Problems Categories
Liara Aparecida dos Santos Leal1 , Dalcidio Moraes Claudio1 ,

Laira Vieira Toscani2 , and Paulo Blauth Menezes2
1
Mathematics Department
PUCRS – Porto Alegre, Brazil
{liara, dalcidio}@pucrs.br
http://www.mat.pucrs.br/
2
Computing Institute
UFRGS – Porto Alegre, Brazil
{blauth, laira}@inf.ufrgs.br
http://www.inf.ufrgs.br/
Extended Abstract
The notion of approximation problems was formally introduced by Johnson [4] in

his pioneering paper on the approximation of combinatorial optimization prob-
lems, and it was also suggested a possible classification of optimization problems
on grounds of their approximability properties. Since then, it was clear that,
even though the decision versions of most NP-hard optimization problems are
polynomial-time reducible to each other, they do not share the same approxima-
bility properties. In spite of some remarkable attempts, according to Ausiello
[1] the reasons that a problem is approximable or nonapproximable are still un-
known. The different behaviour of NP-hard optimization problems with respect
to their approximability properties is captured by means of the definition of ap-
proximation classes and, under the “P 6= NP” conjecture, these classes form a
strict hierarchy whose levels correspond to different degrees of approximation.
In this paper we continue along the same line of research started in [5, 6],
towards to providing a categorical view of structural complexity to optimization
problems. The main aim is to provide a universal language for supporting for-
malisms to specify the hierarchy approximation system for an abstract NP-hard
optimization problem, in a general sense. From the observation that, intuitively,
there are many connections among categorical concepts and structural complex-
ity notions, we started defining two categories: the OPTS category of polynomial
time soluble optimization problems, which morphisms are reductions, and the
OPT category of optimization problems, having approximation-preserving re-
ductions as morphisms. In this direction a comparison mechanism between the
OPTS and OPT categories has been introduced in [7] and, in order to establish
a formal ground for the study of the approximation properties of optimization
problems, a system approximation to each optimization problem is constructed,
based on categorical shape theory [2] and previous works by C. Rattray [9, 10].
Given a functor K: OPTS−→ OPT, the category AP XB,K of approximations to
an optimization problem B ∈ OPT is the comma category B ↓ K of K-objects
under B. A such kind of limit construction provides a means of forming complex
objects from patterns (diagrams) of simpler objects. In particular, by using co-
limits in the AP XB,K definition, a hierarchical structure can be imposed upon
the system of approximation, reaching the best approximation from the system,
if it exists.
Right now, we consider the new insights and a deeper understanding of the
central questions and their implications. The direction is aimed towards ac-
tually exploring the connections among the structural complexity aspects and
categorical concepts. Under this approach, and under reasonable conditions, a
new characterization to approximation problems by means functorial construc-
tion on their approximation categories are still being investigated and left for
further works.
Acknowledgments
This work was partially supported by: PUCRS – Pontificia Universidade Catolica
do Rio Grande do Sul, Brazil, UFRGS – Universidade Federal do Rio Grande
do Sul, Brazil and CNPq – Conselho Nacional de Desenvolvimento Cientifico e
Tecnologico, Brazil.
References
1. G. Ausiello, P. Crescenzi, G. Gambosi, V. Kann, A. Marchetti-Spaccamela and M.
Protasi, “ Complexity and Approximation – Combinatorial Optimization Problems
and their Approximability Properties”, Springer-Verlag, 1999.
2. J-M. Cordier and T. Porter. Shape Theory: Categorical Approximations Methods.
Ellis Horwood Ltd, 1990.
3. M. R. Garey and D. S. Johnson, “Computers and Intractability - A guide to the
Theory of NP-Completeness”, Bell Laboratories Murray Hill, New Jersey, 1979.
4. D. S. Johnson, “Approximation Algorithms for Combinatorial Problems”, J. Com-
put. System Sci. 9, 1974. p.256-278.
5. L. A. S. Leal; P. B. Menezes; D. M. Claudio and L. V. Toscani. Categorias dos Prob-
lemas de Otimização. In: XIV SBES - Workshop on Formal Methods - WFM2000,
3. Oct. 04-06, João Pessoa, Paraı́ba. Proceedings... Brazil: SBC, 2000. p.104 – 109.
6. L. A. S. Leal; P. B. Menezes; D. M. Claudio and L. V. Toscani. Optimization Prob-
lems Categories. In: Formal Methods and Tools for Computer Science - EURO-
CAST’2001. Las Palmas de Gran Canaria, Canary Islands, Spain, R.Moreno-Diáz
and A.Quesada-Arencibia Eds., 2001. p.93 – 96,
7. L. A. S. Leal; P. B. Menezes; D. M. Claudio and L. V. Toscani. Optimization
Problems Categories. EUROCAST’01, LNCS 2178. Springer-Verlag, R.Moreno-
Diáz, B. Buchberger and J-L. Freire Eds., 2001. p.285 – 299.
8. L. A. S. Leal; D. M. Claudio; P. B. Menezes and L. V. Toscani. Modelling the
Approximation Hierarchy to Optimisation Problems Through Category Theory.
International Journal of Computing Anticipatory Systems, v. 11, Belgium, 2002.
p.336 – 349.
9. C. Rattray. Identification and Recognition Through Shape in Complex Systems.
LNCS 1030, Springer-Verlag, 1995. p.19-29.
10. C. Rattray. Abstract Modelling Complex Systems. Advances in Computing Sci-
ence, Systems: Theory and Practice. R. Albrecht, Ed., 1998. pp. 1-12.
Computation of Partial Automata through Span
Composition?
Karina Girardi Roggia, Marnes Augusto Hoff, and Paulo Blauth Menezes
Instituto de Informática - UFRGS

Av. Bento Gonçalves, 9500 - Campus do Vale - Bloco IV
Caixa Postal 15064
Porto Alegre – RS – Brazil
91501-970
{kaqui,marnes,blauth}@inf.ufrgs.br
1 Extended Abstract
In Computer Science, to express computations that don’t terminate and to define

partial recursive functions, it is common to use the notion of partiality. Actually,
due to partiality, p.g., the class of partial recursive functions becomes equivalent
to Turing Machines. Category Theory is becoming an useful tool to formalize
abstract concepts making easy to construct proofs and investigate properties in
many areas, specially in Semantics and Type Theory. The constructions about
universal mappings like limits and adjunctions are getting useful interpretations
in terms of compositionality of systems. Besides, in Category Theory there is
tools to define structures more complex based in simple ones like Comma Cate-
gories. With Comma Categories is possible to define a category based in another
category inheriting properties. Categories of Graphs and Automata are usually
defined by this structure.
In Category Theory, partial morphisms can be seen as a generalization of
the notion of partial functions. Categories with partial morphisms have gain
interest as a tool for theoretical modeling of several aspects of Computation,
for example in Recursion Theory and Universal Algebra, just to name a few.
Robinson and Rosolini [1] describe some approximations to formalize the concept
of categories with partial morphisms. And as pointed out by [2], in Computation
Theory, as well as in actual programming and semantics of programs, diverging
computations cannot be avoided unless a restriction is made to a subclass of the
computable functions.
Graphs are commonly used to model systems, either by simple graphs or by
graph-based structures like Petri nets [3–5] and automata [6, 7]. Automata is
a common formalism used in many areas in Computer Science like compilers,
microelectronics, etc. Most of the study about it is in the Formal Language area.
In this paper we define a different category of automata: a category of Partial
Automata, named Autp . This category is constructed over a category of partial
graphs (Grp ). The difference between a graph and a partial graph is in the
?
This work is partially supported by CAPES and CNPq.
definitions of the source and target functions that mapped an arc to a node:
in graphs these functions are total while in partial graphs source and target
functions are partial functions. Due to this difference, automata based in partial
graphs can have marks of initial and final states naturally. Beyond that, we
can also define constructions like limits in Autp that allows compose partial
automata.
Span Composition [8] can be used to compose graphs like in [9] where it gives
semantics of systems with dynamic topology, p.g.. This kind of composition could
be used to define the computations of automata. Briefly, a span composition of
two partial automata results in a partial automata where each edge represents
a path of length two (between nodes), which first half is some edge of the first
automaton and which second half is some edge of the second one. It is possible
to compose the same automaton with itself several times which is the purpose
of this paper. In the case of n successive span compositions, we can obtain all
the words of its accepted language whose needs n + 1 steps of computation in
the arcs of the partial automaton that don’t have source neither target.
Let a partial automaton that computes the language L. Composing itself ad
infinitum by span composition, the resulting arcs without source neither target
nodes can be seen as the transitive closure L+ . If we compose itself n times, this
kind of arcs will be the subset of L+ whose word’s length is limited to n + 1,
i.e., the computations of the automaton with n + 1 steps.
References
1. Robinson, E., Rosolini, G.: Categories of partial maps. Information and Computa-
tion 79 (1988) 95–130
2. Asperti, A., Longo, G.: Categories, Types, and Structures - An Introduction to
Category Theory for the Working Computer Scientist. MIT Press, Cambridge,
USA (1991)
3. Peterson, J.L.: Petri Net Theory and the Modelling of Systems. Prentice-Hall,
Englewoods Cliffs, New Jersey (1981)
4. Meseguer, J., Montanari, U.: Petri nets are monoids. Information and Computation
88 (1990) 105–155
5. Menezes, P.B.: Diagonal compositionality of partial petri nets. In: 2nd US-Brazil
Joint Workshops on the Formal Foundations of Software Systems, Eletronic Notes
in Theoretical Computer Science v.14 (1998)
6. Hopcroft, J.E.: Introduction to automata theory, languages and computation.
Addison-Wesley (1979)
7. Adamek, J., Trnkova, V.: Automata and Algebras in Categories. 1 edn. Kluwer,
Dordrecht (1990)
8. Bénabou, J.: Introduction to bicategories. In: Reports of the Midwest Category
Seminar. Number 47 in Springer Lecture Notes in Mathematics. Springer-Verlag
(1967) 1–77
9. Hoff, M.A., Roggia, K.G., Menezes, P.B.: Composition of Transformations: A Frame-
work for Systems with Dynamic Topology. International Journal of Computing
Anticipatory Systems 14 (2004) 259–270
Neural network sensitivity analysis applied for
the estimation of sensor matrix redundancy
Przemyslaw M. Szecówka1 , Andrzej Szczurek2 , Benedykt W. Licznerski1 ,

Franz Pichler3
1
Faculty of Microsystem Electronics and Photonics, Wroclaw University
of Technology, Janiszewskiego 11/17, 50-372 Wroclaw, Poland
2
Institute of Environmental Protection Engineering, Wroclaw University
of Technology, Wybrzeze Wyspianskiego 27, 50-370 Wroclaw, Poland
3
Institut fur Systemwissenschaften, Johannes Kepler University,
Altenbergerstr. 69, A-4040 Linz, Austria
Introduction
Contemporary sensor systems consist of the matrices of sensing elements and

smart data processing techniques applied for their responses analysis, provid-
ing the desired information. Typical solutions of gas sensor systems, e.g. [1-7],
although very successful, suffer from the lack of the reasonable method of sen-
sors selection. Eventually most of the systems rely on more or less redundant
set of sensing elements. The solution may be neural network sensitivity analysis
method adopted for the estimation of the value of the information given to the
system by the particular sensors in the matrix. The original method [8] involves
calculation of the neural network output derivative on the selected input. It was
generalized to work with the neural networks containing more hidden layers.
The problems with data scaling and normalisation were considered leading to
the improvement of the reliability of the method.
Experimental
The concept is illustrated by the acetone/chloroform mixtures quantitative analy-

sis using TGS sensors [9]. The matrix initially containing 6 sensors was placed
in a test chamber with controlled atmosphere [11]. The set of vectors containing
the gas concentrations, humidity level and sensors responses was collected this
way. The data was applied to train a series of the neural networks, with the
sensors responses as the input and gas concentrations as the output. The sensi-
tivities of all outputs for all inputs were calculated for all the available samples.
The huge amount of data was reduced using the Euclidean formula. Eventually
the single factor, estimating the impact to the system response was obtained for
each sensor. The sensor with the lowest value of the factor was removed from the
matrix and the whole process was repeated again for the reduced set of sensors.
The first two stages of the experiment, repeated for various structures of neural
networks with 1 and 2 hidden layers brought coherent results, i.e. pointed to
the same sensor as the most redundant one. In the third stage there were some
doubts and in the fourth one (with only 3 sensors left), the order of sensitivity
factors calculated for the series of neural networks trained was different, dis-
abling further reduction of the sensor array. Eventually the reduced sensor array
was obtained, which provided the accuracy of gas concentration measurement
comparable with the initial one at prospectively lower cost.
Conclusions
It was shown that the neural network sensitivity analysis may be attractive tool
for the optimization of sensor systems. It may be considered in the context of
the sensing elements selection and their working conditions setup (e.g. operation
temperature). Similar approach may be also used in the other fields requiring
some estimation of the quality of the series of information sources.
References
1. Moore S.W., Gardner J.W., Hines E.L., Gopel W., Weimar U.: A modified multilayer
perceptron model for gas mixture analysis, Sensors and Actuators B 44: 517-520,
1997.
2. Huyberechts G., Szecówka P.M., Roggen J., Licznerski B.W.: Simultaneous quan-
tification of carbon monoxide and methane in humid air using a sensor array and
artificial neural network, Sensors and Actuators B 45: 123-130, 1997.
3. Di Natale C., Davide F.A.M., D’Amico A., et al.: A composed neural network for
the recognition of gas mixtures, Sensors and Actuators B 24-25: 808-812, 1995.
4. Jervis B.W., Desfieux J., Jimenez J., et al.: Quantification of gas concentrations in
mixtures of known gases using an array of different tin-oxide sensors, IEE P-Sci.
Meas. Tech. 150 (3): 97-106, May 2003.
5. Zou X.B., Zhao J.W., Wu S.Y., et al.: Vinegar classification based on feature ex-
traction and selection from tin oxide gas sensor array data, Sensors 3 (4): 101-109,
April 2003.
6. Martin M.A., Santos J.P., Agapito J.A.: Application of artificial neural networks to
calculate the partial gas concentrations in a mixture, Sensors and Actuators B 77:
468-471, 2001.
7. Sobański T., Saffarpour N.: Analysis of gas mixture using sensor array and artificial
neural networks. In: Cybernetics and systems 2002. Proceedings of the Sixteenth
European Meeting on Cybernetics and Systems Research. Vol. 1.: 97-100, Vienna
2-5 April 2002.
8. Żurada J.M., Malinowski A., Usui S.: Perturbation method for deleting redundant
inputs of perceptron networks, Neurocomputing 14: 177-193, 1997.
9. Figaro Gas Sensors, TGS 800 series specification, Figaro Engineering Inc. 1997.
10. Roussel S., Forsberg G., Steinmetz V., P. Grenier P., Bellon-Maurel V.: Optimi-
sation of electronic Nose Measurements. Part I: Methodology of Output Feature
Selection, Journal of Food Engineering 37: 207-222, 1998.
11. Szczurek A., Szecówka P.M., Licznerski B.W.: Application of sensor array and
neural networks for quantification of organic solvents vapours in air, Sensors and
Actuators B 58: 427-432, 1999.
Fuzzy Modeling for Coal Seams
A Case Study for a Hard-Coal Mine
Luis Argüelles, Jose M. Rivas, Javier Toraño, Carlos González, R. García, Teresa
de Pedro, J. Antonio Martín
Carbonar S.A. Cardenal Cienfuegos 8, 33007 Oviedo

ETSIMO. Independencia, 13, 33004, Oviedo
Industrial Automation Institute, CSIC, La Poveda, Arganda del Rey.
28500 Madrid, Spain
{arguelles, jmrivas}@carbonar.es
jta@uniovi.es
{jamartin, ricardo, tere, gonzalez}@iai.csic.es
Abstract: This work presents a set of linguistic variables for modeling some
geologic and morphological features of a coal seam. Since self-advancing
winning coal systems and their interactions with coal faces represent extremely
complex situations where traditional mathematical models are unable to offer
working solutions, we explore the viability of using Fuzzy-logic based
techniques in order to obtain a good-enough model that allows technicians at
mine to get a better understanding of the variables at play and to foresee the
coal production.
Introduction
Coal mining is a strategic industrial activity in the North of Spain. An example of an

especially competitive mine is that of Carbonar S.A. because of its hard coal seam
about 4m in thickness and the planning and use of modern techniques for coal
winning. Coal thickness up to 6 meters and other physical characteristics of the seam
permits the exploitation with a longwall mining method. This method consists on a set
of self-advancing shields (left part of figure) that are located adjacent to one another
along a transversal section of the seam. The shields are used to keep the hanging-wall
in place while the coal winning machine takes 60 cm off the wall (right part of figure)
for every coal winning run. After the shearer passes, the coal is taken off the mine
with a continuous system of transport (panzers, conveyor belts). A manually operated
mechanism permits the operators to advance the shields.
A research project 1 has been awarded in 2004 to Carbonar S. A. by the “Centro para
el Desarrollo Tecnológico Industrial” (CDTI) agency of the Spanish Ministry of
Industry for financially supporting the research. The goals of the project are to
analyze, model and simulate the longwall exploitation system merging finite element
1 “Sistema de Simulación Inteligente en Arranque Mecanizado Integral”, CDTI 20040116

modeling (FEM), fuzzy logic techniques (FL) and virtual reality (VR). In this paper
we will present the preliminary works of that project. Aside the R+D department of
Carbonar S.A., CSIC’s Instituto de Automática Industrial and the Mining and
Exploration Department of the Mining Engineering School of the University of
Oviedo cooperate in the project.
As exposed, one of the goals of the project is to simulate the behavior of the longwall
equipment with fuzzy logic. Fuzzy logic is a technique whose power to perform
control of systems where the model is not well known has already been proven. In
this work we attempt to use fuzzy logic not to control the longwall, but to model its
interplay with the existing geomorphologic conditions. In order to do this, the project
is planned in three phases:
Left fig. Shield Support Right fig. Coal-winning shearer working.
1. Simulation of the coal winning system. There are many variables influencing
the coal winning activities. This part of the project will attempt to identify
them and assign linguistic values to them. Several output variables will be
established as, for instance, the number of shearer passes per labor shift, or
the daily advance in longitude for the coal seam resulting from winning the
coal. After that, it will be necessary to establish the rules linking the input
and output variables.
2. If problems are detected in the simulation by the fuzzy module under
definite conditions, a detailed physical analysis of the shield mechanism will
be run for these conditions by means of finite elements modelling
techniques.
3. A virtual reality system will show the results of the simulation.
References
1. “Fuzzy Logic in Geology”, Eds. Rober V. Demicco, George J. Klir, Elsevier Academic
Press, 2004 , ISBN 0-12-415146-9.
Data Analysis of a Distributed Solar Collector
Field
Manuel Berenguel2 , Ryszard Klempous1 , Henryk Maciejewski1 , Jan Nikodem1 ,

Maciej Nikodem1 , Loreto Valenzuela3
No Institute Given
Extended abstract
The purpose of this work is to analyze methods, and algorithms to control the
ACUREX solar collectors field of the Platforma Solar de Almerı́a - PSA. The
analysis of me6thods and algorithms to control the PSA solar field is a starting
point to formulate a model of solar collectors field control optimization. The sole
criterion of our optimization policy is to maximize the energy accumulated in the
oil storage tank but the optimal strategy should fulfill some constrains resulting
from physical and technical restrictions. It results in a nonlinear dynamic opti-
mization problem with nonlinear constraints, which is considered in literature
as hard. The data concerned was related to performance of various control algo-
rithms observed under changing environmental conditions. The main concepts
of the approach discussed are:
1. Prior to data analysis tasks, data preprocessing stage takes place, the pur-
pose of which is to:
– integrate data from different sources e.g., combine data from different
monitoring devicessystems into one consistent data model,
– ensure good quality of data for analysis (preliminary analysis of data
to assess data quality: discover data integrity problems, ’outliers’, miss-
ing values, etc.; design and implement data transformations to resolve
discovered quality problems).
2. Integrated data is stored in a data warehouse, which is a dedicated repository

of data for analytical processing. Data warehouse implements a logical data
model convenient for analysis.
3. Data analysis is performed using OLAP tools, which give users intuitive
interface (usually Webbased) to formulate data analysis tasks.
4. Data mining algorithms and tools can be subsequently applied which allow
analyst to discover relationships and patterns in data.
The real power of the approach proposed lies in the application of:
– warehouse and OLAP techniques

– software package for statistical data analysispresentation (in the project dis-
cussed the SAS System was used) to the task of explanatory data analysis
from control system monitoring.
References
1. ACUREX Field D.A.S. and Control System, Documentation version 2001,Platforma
Solar de Almerı́a, Almerı́a, Spain 2001.
2. Blanco-Muriel M., Alarcón-Padilla D.C., López-Moratalla T., Lara-Coira M.: Com-
puting the solar vector, Solar Energy 70(5):431-441, 2001.
3. Camacho E.F., Rubio F.R., Hughes M.: Self-tuning control of a solar power plant
with a distributed collector Field, IEEE Control Systems Magazine, 72-78, 1992.
4. Camacho E.F., Berenguel M., Rubio F.R.: Advanced control of solar plants, Ad-
vances in Industrial Control, Springer-Verlag, 1997.
5. Carmona R.: Analisis, modelado y control de un campo de colectores solares dis-
tribuidos con sistema de seguimiento en un eje, PhD thesis, Universidad de Sevilla,
1985.
6. Dorf R., Bishop R.: A modern control systems, Addison-Wesley Longman Inc.,
Menlo Park CA, 1998.
7. Farkas I., Vajk I.: Experiments with internal model-based controller for Acurex
Field, IHP Programme, Proceedings of the 2-nd Users Workshop, Almerı́a, Spain,
2002.
8. Johansen T., Storaa C.: An internal controller for distributed solar collector fields,
IHP Programme, Proceedings of the 2-nd Users Workshop, Almerı́a, Spain, 2002.
9. Pereira C., Dourado A.: Application of a neuro-fuzzy network with support vec-
tor learning to a solar power plant, IHP Program, Proceedings of the 2nd Users
Workshop, Almerı́a, Spain, 2002.
10. Silva R., Rato M., Lemos J.: Observer based time wrapped control of distrib-
uted collector solar fields, IHP Program, Proceedings of the 2nd Users Workshop,
Almerı́a, Spain, 2002.
11. Smolec W.: Fototermiczna konwersja energii sÃlonecznej, Wydawnictwo Naukowe
PWN, Warszawa, 2000.
12. Solar thermal electricity generation, Lectures from the summer school at the P.S.A.
13-17 July 1998, Almerı́a, Spain.
13. Valenzuela L., Balsa P.: A series and parallel feedfoward control schemes to regulate
the operation of a solar collector field, Serie Ponencias, Proceedings of the 1st Users
14. Vaz F., Oliviera R.: Neves de Silva R.: PID control of a solar plant with gain
interpolation. Serie Ponencias, Proceedings of the 1st Users Workshop, Almerı́a,
Spain, 1997.
15. Zarza E. at all: The DISS Project: Direct Steam Generation in Parabolic Trough
Systems. Operation and Maintenance Experience and Update on Project Status.
Journal of Solar Energy Enginnering Transactions of the ASME, 124(2):126- 133,
2002.
Non-linear Formal Modelling
of Amphibian Tectal Cells
Gabriel de Blasio, Roberto Moreno-Diaz
Instituto Universitario de Ciencias y Tecnologı́as Cibernéticas

Universidad de Las Palmas de Gran Canaria
gdeblasio@dis.ulpgc.es
rmoreno@ciber.ulpgc.es
Extended Abstract
The complexity of modelling and simulation of cells in the visual pathway

varies from lower to higher vertebrates. At each stage (retina, optical ’ceiling’
(tectum), or lateral geniculate) the nature of the inputs must be established
before any attempt in modelling. This can be relatively straightforward for higher
vertebrates, because, apparently , there is little processing in foveal parts of the
retinal signals as they travel from the retina to the visual cortex. For amphibians
there is a different situation because highly non-linear specialized processing
takes place at the level of retinal ganglia. Amphibians have no visual cortex,
and the structure of the higher visual processing station, the optical tectum,
seems to be related directly with motor behavior for procuring food or escaping
from enemies. Other vital activities (e.g. mating) rely on other senses.
To proceed in modelling and simulating the tectal cells of the amphibia ac-
cording systems theory concepts it is first required to identify the nature of the
input and output relations of signals. The problem is then to propose plausible
neuronal mechanisms, in agreement with neurophysiology and neuroanatomy,
to account for the input-output relations under different circumstances and to
predict results in new experimental situations.
Inputs to the tectal system are the ganglion retinal cells outputs carried by
their axons. After the optic chiasm, the axons project in different layers of the
tectum, according to the nature of the computing ganglion cell.
For the frog (Rana Pipiens), the four types of registered ganglia outputs at
different (four) layers of tectum are
– 1. Group 1 ganglion cells respond to edges, i. e. , to differences in illumination

on the receptive field.
– 2. Group 2 ganglion cells respond to the image of a small dark object, moving
centripetally into its visual field.
– 3. Group 3 ganglion cells respond to any change, spatial or temporal, in the
luminance of the scene.
– 4. Group 4 ganglion cells respond to dimming of light.
Lettvin, et al. were the first to find correlations between the form and function
of each of these cells [1].
Their description of the tectum was focused only to the response of tectal
cells in a prey-catching behavioral situation.
In Pedro Ramon y Cajal’s drawing of the frog’s tectum there are four main
features that appear clearly. First, the cell bodies lie in the deeper layers . Sec-
ond, the neurons seem to have a restricted dendritic tree. Third, axons leave
the tectum branching from a main ascending dendrite. Fourth, interconnections
among cells lying at the same depth is strongly suggested.
Fig. 1. (a) Schematic diagram of frog’s tectum. (b) Diagram of tectal cell interaction
processes
Figure 1(a) shows what appear to be the essential anatomical features of

the tectum. Axons from the four major retinal ganglion cell groups enter the
tectum as the optic tract and map onto the four superior tectal layers called
the superficial neuropile. Axons from gang-lion cells which respond to adjacent
retinal areas map onto adjacent points in the neuropile. Thus, each small area
in the retina corresponds to four small areas, one on each layer, in the neuropile.
Within the neuropile there also exist a few cells, not represented in the figure,
which appear to be lateral connectives.
All other tectal cell bodies lie in the granular layer, beneath the superficial
neuropile. Their dendrites, which have the appearance of columns, extend up
through the four neuropile layers covering a small area in each layer. Their
axons leave the tectum by branching from the main dendrites, at the lower part
of the neuropile.
For toads (Bufo Bufo), the most extensive neurophysiological recordings, and
the correlation of these with behavioral patterns have been made by Ewert [2],[3]
also in prey-catching situations.
In the toad retina there are similarly three types of ganglion cells that send
their fibers by way of the optic nerve to the optic tectum. Groups 1 is missing.
The three ganglion-cell types in the toad (as in the frog) differ in several charac-
teristics, including in particular the diameter of their excitatory receptive fields:
about four degrees for the so-called Group 2 ganglion cells, about eight degrees
for Group 3 cells and from 12 to 15 degrees for Group 4 cells.
In what respect to tectal cells, classical studies [4] by Lettvin et al. have dis-
tinguished from several varieties of tectal neurons two extreme types: ”Newness”
cells, N, sensitive to new objects in the visual field, and the ”Sameness” cells. S,
sensitive to the same object for a specific period of time. The bodies of Sameness
cells lie deep in the granular zone. The properties of Newness and Sameness cells
are shown by the following outline:
– a. Newness Cell
– 1. Its receptive field is approximately 30◦ in diameter, with considerable
overlap.
– 2. It yields small responses to sharp changes in illumination.
– 3. Its response frequency increases if the motion of an object is irregular.
Its response also depends on the velocity and size of the object and on the
direction of motion.
– 4. It has no enduring response.
– 5. Its adaptation time is less than 1 second.
– 6. Adaptation is erased with a step to darkness.
– b. Sameness Cell
– 1. Its receptive field is almost all the visual field, but includes a null region.
– 2. It does not respond to changes in illumination.
– 3. It responds to the presence of an object in its field by generating a low
frequency pulse train. Response is maximum to objects about 3◦ in diameter.
– 4. It ”notices” with a ”burst” of pulses all motion discontinuities. It dis-
criminates among objects, fixing its ”attention” on the one with the most
irregular movement.
Lettvin, et al.comment that N and S cells may be considered two extremes of

several types of tectal cells. From Pedro Ramon y Cajal’s drawings, tectal cells
do not seem to be significantly different, anatomically, although S cells could
receive excitation from N cells. A general tectal model could be en-visioned,
in which N and S cells would be differentiated only by values of parameters,
such as adaptation time, width of the receptive field, etc. This is the point of
view adopted here. While the receptive field is relatively wide for both types of
cells, the dendritic tree is very restricted. Interaction between tectal cells would
account for this observation.
For toads, Ewert also distinguishes two essential types, I and II, which may
correspond to Lettvin’s N and S tectal cells.
The integration of all this experimental data leads to a formal model which
is represented in figure 1(b), and which is a generalization by using more modern
tools [5], of the first proposal or a model of frogs tectum [6].
In the conceptual model of tectal-cell behavior, object velocity and size is
determined by signals assumed to originate on the retina. For convenience these
current and/or voltage signals are referred to simply as ’activity’. Assume that
each layer of the superficial neuropile is divided into small areas, such as area K
in layer i(i = 1, 2, 3, 4). Some ganglion-cell axons of Group i map into this area
and emit their signals there. Assume a decoding process which produces analog
values such as voltage signals. Let ViK (t) be the voltage signal indicating the
response of the ganglion cells which map into area K of layer i.
Assume further that the dendrites of each tectal neuron are, restricted to
an area K i. e. there is no overlapping. The first process that is suggested is a
summation of the signals, ViK (t), for the same K and different i’s at time t.
The corresponding ViK (t) can be excitatory or inhibitory. By summation
a series of ’vertical’ lines is obtained, which transmit signals VK (t) through
the superficial neuropile. These signals interact due to the conductivity of the
medium through which they pass. Four interaction processes account for tectal
cell properties. Process A at the level of the dendrites, and Process D, at the
level of the axons, are of the same nature, i. e. , an interaction due to the
conductivity of the medium. Process B, which accounts for adaptation, is similar
to a distributed capacity which, once charged, damps the transmission of the
signals. Process C requires non linear operations that could be provided by the
cellular membrane.
The main properties of tectal cells are accounted for in a simple manner by the
formal model. The main hypothesis is that in essence the same basic mechanism
should be assumed for both Types N and S (I and II) tectal cells. To produce
the desired response for the cells, this mechanism should include at least four
neurophysiologically plausible processes namely, lateral facilitation, adaptation,
selection of maximum activity, distribution of it, and probably feedback from
the subtectum.
References
1. Lettvin, J.Y. et al: What the Frog’s Eye Tells the Frog’s Brain. Proc. I.R.E., 47
(1959) 1940–1951
2. Ewert, J.P.: Neuroethology of Releasing Mechanism: Prey-Catching in Toads. Be-
havioral and Brain Sciences, 10 (1987) 337–405
3. Ewert, J.P.: The Neural Basis of Visually Guided Behavior. Scientific American,
230(3) (1974) 34–42
4. Lettvin, J.Y. et al: Two remarks on the Visual System of the Frog. in Sensory
Communication. W. A. Rosemblith ed. MIT Press, (1961) 757–776
5. Moreno-Dı́az, R., de Blasio, G.: Systems Methods in Visual Modelling. Systems
Analysis Modelling Simulation 43(9) (2003) 1159–1171
6. Moreno-Dı́az, R.: Conceptual Model of the Frog’s Tectum. Sensory, Decision and
Control Systems. R-548. MIT, MA. USA, (1967)
Optimization of a Class of Uncertain Systems Based on
Uncertain Variables
Z. Bubnicki
Institute of Control and Systems Engineering, Wroclaw University of Technology

Wyb. Wyspianskiego 27, 50-370 Wroclaw, POLAND
zdzislaw.bubnicki@pwr.wroc.pl
Extended Abstract
The idea of uncertain variables has been developed for a class of uncertain systems
described by a relational knowledge representation [1], [2], [3]. This paper presents a
new result in this area: an optimization problem formulated by using the uncertain
variables. The problem formulation is described after a short presentation of the
uncertain variables and a basic decision problem.
In the definition of the uncertain variable x we consider two soft properties:
~
" x ≅ x" ( x is approximately equal to x) and " x ∈ D " ( x approximately belongs to
D, i.e. its approximate value belongs to D) where D ⊆ X and X is a set of values
(vector space). The uncertain variable is defined by a set X and a certainty distribution
h ( x) = v ( x ≅ x) given by an expert, where v ∈ [0,1] denotes the certainty index of the
~
soft property, v ( x ∈ D) = max h ( x) . In the version called C-uncertain variable we
x∈ D
use the certainty index
~ 1
vc ( x ∈ D) = [max h ( x ) + 1 − max h ( x)].
2 x∈ D x∈ X − D
Let us consider a static plant with the input vector u ∈ U and the output vector
y ∈ Y , described by a relation R (u , y; x) ⊂ U × Y where the vector of unknown
parameters x ∈ X is assumed to be a value of an uncertain variable described by the
certainty distribution h (x) given by an expert. If the relation R is not a function then
the value u determines a set of possible outputs D y (u; x) = { y ∈ Y : (u , y ) ∈ R} . For
the requirement y ∈ D y ⊂ Y given by a user, we can formulate the following
decision problem: For the given R (u , y; x ) , h (x ) and D y one should find the
decision u * maximizing the certainty index that the set of possible outputs
approximately belongs to D y (i.e. belongs to D y for an approximate value of x ).
Then
~
u * = arg max v [ D y (u; x ) ⊆ D y ] = arg max max h ( x) (1)
u ∈U u ∈U x∈ Dx (u )
where D x (u ) = {x ∈ X : D y (u; x ) ⊆ D y } . The considerations for C-uncertain variable
are analogous, with v c instead of v .
The purpose of this paper is to present a formulation and solution of a specific
optimization problem for the uncertain plant under consideration. The main idea is as
follows: Consider the plant described by R(u, y; x) where one-dimensional y
denotes a quality index which should be minimized. In this case, the requirement may
be presented in the form y ≤ α or D y = (−∞, α ] where α is given by a user and
should be as small as possible. Consequently, the decision u * (α ) determined by (1)

∆
and the corresponding certainty index v [u * (α )] = v (α ) are the functions of α . It is
easy to note that, as a rule, v (α ) is an increasing function. For the given threshold v
determining the required level of the uncertainty, one should solve the equation
v (α ) = v with respect to α , and use the solution α in the determination of the final
results: u* = u* (α ) and v* = v (α ) . The corresponding algorithm, the scheme of the
decision making system in this case, examples and results of simulations are
presented for two versions: based on the basic uncertain variables with v and on C-
uncertain variables with vc .
It has been shown how to extend the presented approach to the optimization for a
class of uncertain dynamical systems described by a set of relations or by a classical
functional model with unknown parameters which are assumed to be values of
uncertain variables described by certainty distributions given by an expert. The
further extension is possible for uncertain systems with uncertain and random
parameters [4], [5], and with a distributed knowledge [6].
References
1. Bubnicki, Z.: Uncertain Variables and Their Application to Decision Making. IEEE Trans.
on SMC, Part A: Systems and Humans 6 (2001) 587–596
2. Bubnicki, Z.: Uncertain Logics, Variables and Systems. Springer-Verlag, Berlin London
New York (2002)
3. Bubnicki, Z.: Analysis and Decision Making in Uncertain Systems. Springer-Verlag, Berlin
London New York (2004)
4. Bubnicki, Z.: Stability and Stabilization of Discrete Systems with Random and Uncertain
Parameters. Proceedings of the 15th IFAC World Congress, Vol. E “Robust control”.
Pergamon, Oxford (2003) 193–198
5. Bubnicki, Z.: Application of Uncertain Variables in a Class of Control Systems with
Uncertain and Random Parameters. Proceedings of the European Control Conference,
Cambridge, UK (2003)
6. Bubnicki, Z.: Application of Uncertain Variables to Decision Making in a Class of
Distributed Computer Systems. In: Musen, M., Neumann, B., and Studer, R. (eds.):
Proceedings of 17th IFIP World Computer Congress. Intelligent Information Processing.
Kluwer Academic Publishers, Norwell MA (2002) 261–264
Computational Simulation of Categorical
Constructions?
Rodrigo Born Vieira and Paulo Blauth Menezes
Computing Institute
UFRGS - Porto Alegre, Brazil
{rodrigor,blauth}@inf.ufrgs.br
http://www.inf.ufrgs.br/
1 Extended Abstract1
Category Theory is a branch of Pure Mathematics with a scientific field appar-

ently distinct from Computer Science ones. However, some of its properties -
especially expressiveness - make this mathematical model very helpful to the
studies of Computer Science.
Jean Piaget has described in [5] the results he has obtained in experiences
with children. He shows that the idea of composition, typical on categorical
system, is in the genesis of human being. In fact, human naturally gets inclined
to think in schemes simple properties and then, successively, similar operations in
schemes of schemes, and so on, constructing a chain of composition of properties.
Piaget’s results conduct for the conclusion that the intuitive human reasoning
is categorical. This conclusion not only shows that the Category Theory could be
taught earlier than in under graduation or graduation level but also that it should
be done. The development of logic formal thought on children allows developing
skills and reasoning lines which normally aren’t developed on childhood, despite
of being required in many areas of knowledge. This kind of skill and in particular
general and unified though, typical of Category Theory, although being intuitive
for human being, is discouraged beside the learning process.
Category Theory reasoning is used to be learned in some computer science
graduation courses and, less frequently, under graduation courses. Its mathemat-
ical principles, which could be easy, natural and intuitive, become very dense and
abstract because of the way we have been stimulated to reason. Additionally,
the heterogeneity in the presentation of introductory topics and the lack of pub-
lished works in other languages than English complete the set of obstacles to be
surpassed before applying Category Theory in Computer Science.
So, to make this abstract theory become concrete and kind of intuitive is a
challenge and an interesting area for researching. This context has been moti-
vating us to research about computational simulators.
?
This work is partially supported by CAPES and CNPq.
1
Submitted to the workshop Systems Theory and Simulation: formal approaches and
applications
It’s well known that there are limits on representing structures and solving
problems in a computer. Although the computational simulation of Category
Theory is an unexplored area, it’s easy to see that isn’t possible to construct
a simulator able to represent all Category Theory. By deduction, isn’t possible
to implement the verification of a property that is able to work in any kind of
category.
However, the research in computer science shows we can compute and rep-
resent relevant structures and operations, as finite categories with objects being
finite sets and morphisms being relations. This kind of simulator is an efficient
way to turn concrete these abstract concepts, making this theoretical concept
closer to related interests in applied researches.
In spite of the utility for computer science of this kind of simulators, there
is a lack of research in the area. The existent simulators lead only with basic
categorical concepts [2] [4] or are inaccessible to laymen [8].
Watching this scenery, we have been researching about technologies to im-
plement categorical constructions and have been implementing the simulators.
In this paper we present the technologies used to implement these categorical
constructions and the theoretical principles we chose to simulate computation-
ally.
References
[1] Saunders Mac Lane. Categories for the Working Mathematician. Springer Verlag,
New York, USA, 2nd edition, 1998.
[2] R. Rosebrugh and J. Bradbury. Category Theory Database Tools. Mount Allison
University, Canada, 1998. Available at http://cs.mta.ca/research/rosebrugh/ctdt/.
[3] Michael Barr and Charles Wells. Category Theory for Computing Science. Prentice
Hall International Series in Computer Science. Prentice Hall, 2nd edition, 1995.
[4] M. Fleming, R. Gunther, R. Rosebrugh. A Database of Cat-
egories. Mount Allison University, Canada, 1995. Available at
http://mathcs.mta.ca/research/rosebrugh/dbc/.
[5] Jean Piaget. Morphisms and Categories - Comparing and Transforming. Lawrence
Erlbaum Associates, Hillsdale, New Jersey, 1992.
[6] Andrea Asperti and Giuseppe Longo. Categories, Types, and Structures - An Intro-
duction to Category Theory for the Working Computer Scientist. MIT Press, Cam-
bridge, USA, 1991.
[7] Peter J. Freyd and Andre Scedrov. Categories, Allegories. North Holland Publishing
Company, 1990.
[8] D. E. Rydeheard and R. M. Burstall. Computational Category Theory Project.
Electronic edition, 1988. Available at http://www.cs.man.ac.uk/ david/categories/.
Composing Transitions Into Transactions in
UML Diagrams
Júlio Pereira Machado1 and Paulo Blauth Menezes2

1
Faculdade de Informática, Pontifı́cia Universidade Católica do Rio Grande do Sul
Av. Ipiranga 6681, Prédio 30
CEP 90619-900, Porto Alegre, RS, Brasil
juliopm@inf.pucrs.br
2
Instituto de Informática, Universidade Federal do Rio Grande do Sul
Av. Bento Gonçalves 9500, Caixa Postal 15064
CEP 91501-970, Porto Alegre, RS, Brasil
blauth@inf.ufrgs.br
Extended Abstract
The Unified Modeling Language (UML) [1–3] may be used to describe both
the structure and behavior of object-oriented systems using a combination of
notations. For the modeling of the dynamic behavior, a number of different
models are offered such as interaction, state and activity diagrams.
Activity diagrams are one of the means for describing behavior of systems
within the UML language and has been chosen here for its capability in describ-
ing flows of activities of a desired system. Broadly speaking, states of an activity
diagram mostly represent the execution of sequential and concurrent steps in a
computational process. In UML, those steps may be an action or a subactiv-
ity state. Action states represent an atomic computation without substructure
and that cannot be interrupted. Subactivity states, on the other hand, repre-
sent non-atomic execution and can be decomposed and interrupted by external
events.
When modeling concurrent or parallel systems with such diagrams, we must
be aware that basic activities of each system may be constituted by smaller
activities, i.e. transitions are conceptually refined into transactions. Nevertheless,
the UML seems to lack compositional constructs for defining atomic activities.
In this work, we concentrate on describing groups of sequential or concurrent
activities that are responsible for performing a computation, and we address
the issue of modeling transactions. We remark that in our setting the term
“transaction” denotes a certain activity of the system that might be composed by
many, possibly concurrent, subactivities. Moreover, we require this composition
of activities to be considered atomic.
The importance of discussing composition mechanisms is based on the fact,
as pointed in [4], that in order to master the complexity of systems, models
demand a form of structural decomposition. When building large systems, we
need constructs to describe systems as compositions of other smaller systems. In
this sense, when modeling a computational process, we need means of composing
subactivities both in a non atomic or atomic way.
We believe transactions are an important part of today systems and they

deserve a first class mechanism in modeling languages, especially UML. Follow-
ing that premise, this work presents an extension to UML diagrams centered
on a construction for defining atomic composition of actions and activities. We
extend the UML notation with a particular stereotype for composite states (see
figure 1 for a very simple example) and define its semantics by the means of
refinement morphisms over nonsequential automata [5]. The notation is based
on extension mechanisms of UML and introduce new constraints over the basic
composite state, so the new “transaction state” is not allowed to be interrupted;
nonsequential automata constitute a non interleaving semantic domain, with its
foundations on category theory [6], for reactive, communicating and concurrent
systems, following the so-called “Petri nets are monoids” approach [7] and is sim-
ilar to Petri nets, but it is a more concrete model - it can be seen as computations
from a given place/transition net; refinement [8] are formally defined through
a special morphism between automata that maps transitions from the source
automaton (abstract model) to transactions of the target automaton (concrete
model).
Fig. 1. UML activity diagram with transaction composite state

Although UML has a precisely-defined syntax, it still lacks a generally ac-
cepted formal semantics which precisely fixes the meaning of its active behavior
diagrams. Other approaches to translating UML diagrams into formal models
have been based on Petri nets [9]. For example, [10] describes a formal translation
of activity and collaboration diagrams into place/transition Petri nets and [11]
compares different proposals for the semantics based on Petri nets. Also, other
works have used formal methods to verify the behavior of UML specifications
[12, 13].
In this work we further develop the concepts sketched in [14], specially the
translation schemes for building nonsequential automata from UML behavioral
diagrams. Different lines of research are related to this work, manly semantics
based on Abstract Machines, Petri nets and Workflow Statecharts. In [15] ab-
stract state machines are used to provide semantics for activity diagrams in
UML; [16] uses a process-like algebra for describing activity diagrams; and ex-
tensive work has been put on defining semantics for modeling workflow in [11,
17].
References
1. Fowler, M.: UML Distilled: A Brief Guide to the Standard Object Modeling Lan-
guage. 3 edn. Addison-Wesley (2004)
2. Booch, G., Rumbaugh, J., Jacobson, I.: UML User Guide. Addison-Wesley (1999)
3. Group, O.M.: Unified Modeling Language Specification, version 2.0. Rational
Software Corporation (2003)
4. André, C., Rigault, J.P.: Variations on the semantics of graphical models for
reactive systems. In: System Management and Cybernetics, IEEE Press (2002)
5. Menezes, P.B., Sernadas, A.S., Costa, J.F.: Nonsequential automata semantics for
concurrent, object-based language. In: Electronic Notes in Theoretical Computer
Science - 2nd US-Brazil Joint Workshops on the Formal Foundations of Software
Systems. Volume 14., Elsevier (1998)
6. Barr, M., Wells, C.: Category Theory for Computing Science. Prentice Hall (1990)
7. Meseguer, J., Montanari, U.: Petri nets are monoids. Information and Computation
88 (1990) 105–155
8. Menezes, P.B., Costa, J.F., Sernadas, A.S.: Refinement mapping for general (dis-
crete event) system theory. In: Lecture Notes in Computer Science - 5th Interna-
tional Conference on Computer Aided Systems Theory and Technology. Volume
1030., Springer-Verlag (1996) 103–116
9. Reisig, W.: Petri Nets: an introduction. Volume 4 of Eatcs Monographs on Theo-
retical Computer Science. Springer-Verlag (1985)
10. Gehrke, T., Goltz, U., Wehrheim, H.: The dynamic models of UML: Towards a
semantics and its application in the development process. Technical Report 11/98,
Institut fur Informatik, Universitat Hildesheim (1998)
11. Eshuis, R., Wieringa, R.: Comparing petri net and activity diagram variants for
workflow modelling - a quest for reactive petri nets. In: Lecture Notes in Computer
Science - Petri Net Technology for Communication Based Systems. Volume 2472.,
Springer-Verlag (2003) 321–351
12. Shen, W., Compton, K., Huggins, J.: A validation method for uml model based
on abstract state machines. In: Proceedings of EUROCAST. (2001) 220 – 223
13. Knapp, A., Merz, S.: Model checking and code generation for uml state machines
and collaborations. In: Proceedings of 5th Workshop on Tools for System Design
and Verification. (2002) 59 – 64
14. Machado, J.P., Menezes, P.B.: Modeling transactions in uml activity diagrams via
nonsequential automata. In: Actas de la XXX Conferencia Latinoamericana de
Informatica, CLEI (2004) 543 – 553
15. Borger, E., Cavarra, A., Riccobene, E.: An ASM semantics for UML activity
diagrams. In: Lecture Notes in Computer Science - 8th International Conference on
Algebraic Methodology and Software Technology. Volume 1816., Springer-Verlag
(2000) 293–308
16. Rodrigues, R.W.S.: Formalising UML activity diagrams using finite state processes.
In: Workshop on Dynamic Behaviour in UML Models, 3rd International Conference
on the Unified Modeling Language. (2000)
17. Eshuis, R.: Semantics and Verification of UML Activity Diagrams for Workflow
Modelling. PhD thesis, University of Twente (2002)
Acknowledgments
This work was partially supported by CTXML/Microsoft/PUCRS, CNPq (Pro-
jects HoVer-CAM, GRAPHIT, E-Automaton) and FINEP/CNPq (Project Hy-
per-Seed) in Brazil.
One formal systemic approach in software design. Case
study: Extreme programming
Cestmir Halbich1
1Department of information technologies, Faculty of economics and management, Czech

university of agriculture Prague , Kamycka 129, 165 21 Prague – 6, Suchdol, Czech republic
halbich@pef.czu.cz
Abstract. The position paper comes out from Donald Schoen’s ideas about re-
flective practitioners and describes author’s experience in area of computer sci-
ence. The software gap is mentioned and methods for its overbridging too. At
the case study is illustrated Schoen’s approach to designer’s activity and benefit
for improving effectiveness and efficiency of the software design process by re-
flective practitioner’s approach. In the conclusion the advantages and disadvan-
tages are described.
1 Introduction
At the begining of 3rd millennium methods which help develop software products
become more important. They adjust it rapidly and flexibly to changing common
conditions. Often we call these methods agile methodologies. The idea of these is
simple. Only one way exists to check correctness of the developed system - the fastest
development as is possible and then to present results to the customer and adjust it
according by a feedback. Basic principles of agile methods are next:
- incremental and iterative development with very short iterations
- special emphasis on personal communication in frame of development team,
- permanent communication with user
- rigorose automated tests.
From historical point of view we know two approaches to system development. Hard
systems approach is based on systems engineering and systems analysis.The people
are treated as passive observers of the system development process. However, this
approach is not suitable in organizational environment that involves political, social,
or human activities. Development of such systems require an active involvement of of
every stakeholder. The approach that encompasses all the stakeholders of the system
is soft system approach. SSM is a methodology that adopts such ”soft system ap-
proach”.
The first is specific, with clear boundaries and parameters. Possible solutions are
implicit in the problem. The problem could be passed to a designer who could pro-
pose and implement a solution. The second is much less clear. It describes a problem
with weak boundaries and an unclear cause. In fact, it is not describing a problem, so
much as two symptoms. Until the symptoms are explored further, it is not possible to
say what the problem is, let alone propose a solution. Traditional systems analysis is
based on hard problems. These problems have a clear boundary and involve tangible
factors such as staff, documents, procedures, equipment, and structure. Its purpose is
to propose and design a solution for a predetermined problem.
Many business situations are less clear-cut, making a hard approach less appropriate.
Peter Checkland [1] developed Soft Systems Methodology (SSM), recognising that
many problems in organisations are affected by less tangible factors. These include
culture, informal interactions, and attitudes – which he terms the ”Human Activity
System”. SSM’s purpose is to investigate, understand and identify a problem. Its
investigation may reveal a number of problems to address rather than presuppose a
single root cause of difficulties.
SSM is based on systems thinking. It views the problem domain in a holistic rather
than reductionist way, recognising that the component parts are interconnected, so
that a change to one part will affect the other parts. Not only this, but the problem
domain itself is a subsystem of several larger systems – changes in one will affect our
domain as well.
Although Checkland coined the term Soft Systems Methodology, it is not strictly a
methodology. It does not prescribe a rigid series of steps that must be followed scru-
pulously. Rather, it suggests a framework for enquiry, with a number of recom-
mended techniques. The end result is an understanding of the problem domain so that
a hard study can then be applied to specify a solution.
2 One Formal Systemic Approach in Software Design by Extreme

Programming
By Schoen designers use design representations to create a virtual ”design world”[2]

in which the objects and relationships of the design situation are named and made
explicit. The situation that the designer constructs with representations enables new
understandings of the design problem by allowing the designer to see ideas that be-
fore existed only tacitly, and to understand implications of the design situation that
did not exist before constructing the representations. In accordance with some authors
we can say that putting ideas down on paper is not a matter of emptying out the mind
but of actively reconstructing it, then forming new associations, and expressing con-
cepts in pictorial, linguistic, or any explicit representational forms. Designers work
mainly with matter, software developers work usually with algorithms, but the anal-
ogy is evident.
Knowledge is constructed in programming through an interaction between the pro-
grammer's understanding of the design situation and the representations the pro-
grammer’s creates. The software design process is driven by the interaction between
the programmer and design algorithms, rather than by following a pre-planned solu-
tion approach.
Knowledge is constructed in software design through repeated cycles of representing
and interpreting the design situation. Programmers construct the design situation that
talks back in the form of a breakdown. This back talk is interpreted, or reflected upon,
to activate new understandings of the design situation. In this situation is very useful
to use certain formal systemic methodology to improve the effectiveness and effi-
ciency of the software design process.
Already at the conference NATO 1968 the existence of the gap - ”gap, between what
was hoped for from a complex software systems, and what was typically achieved.”
There is a widening gap between ambitions and achievements in software engineer-
ing. ”This gap appears in several dimensions: between promises to users and per-
formance achieved by software, between what seems to be ultimately possible and
what is achievable now and between estimates of software costs and expenditures.
The gap is arising at a time when the consequences of software failure in all its as-
pects are becoming increasingly serious.” [3, p.70]. Since this time amount of differ-
ent methodologies was invented. One of these is so called extreme programming [4].
By author’s opinion the extreme programming is very closely joined with Donald
Schoen’s ideas about reflective practitioners. Contrary to some other programmer’s
techniques extreme programming closely collaborates with users in the all phases of
software design. Some features of extreme programming are discussed below.
Extreme Programming (XP) is actually a deliberate and disciplined approach to soft-
ware development. It has already been proven at many companies of all different
sizes and industries world wide. XP is successful because it stresses customer satis-
faction. The methodology is designed to deliver the software your customer needs
when it is needed. XP empowers your developers to confidently respond to changing
customer requirements, even late in the life cycle. XP implements a simple, yet effec-
tive way to enable groupware style development. XP improves a software project in
four essential ways; communication, simplicity, feedback, and courage. XP pro-
grammers communicate with their customers and fellow programmers. They keep
their design simple and clean. They get feedback by testing their software starting on
day one.
We can visualise the process of the software development by Extreme programming
in a way of scheme which is described in Fig. 1. The scheme is adopted by author
from similar schemes in SSM.
In next case study we can describe software development by XP for very paranoidal
customer. The goal is to develop a call centre support for our customer. But by cus-
tomer’s opinion the project is very secret therefore we don’t know the whole specifi-
cation so yet. The traditional or waterfall methodology is the forefather of all other
methodologies and we find it is most suitable for projects where the requirements are
clearly stated and static or where it helps to have a rigid management structure. The
waterfall methodology is not usable for the development of this information system
because we don’t know the whole specification. The customer defines his demands
only for next two days of the software development. In this case the XP is suitable
because we needn’t analyse the whole system, we develop only small pieces of sys-
tem from day to day. The customer is at our office two days per week and consults
results of our work and his requirements.
problem of real world
- information system
version
of soft-
Action ware
to improve
comparison
Fig. 1. Scheme of the iterations and feedback loops in the software development by Extreme
programming.
3 Conclusion
We have good personal experience with implementing of the extreme programming

as a formal systemic approach in the practice. Reflective practitioner method im-
proves added value of the collaboration in software design process by iterations and
feedback loops with customers and end users. The case study’s Czech company has
good experience with software design of small, medium and large information sys-
tems and work with small and medium work teams.
References
1. Checkland, P.: Systems Thinking, Systems Practice, (1981), John Wiley & Sons, London
(1981)
2. Schoen, D. "Designing as reflective conversation with the materials of a design situation,"
Knowledge-Based Systems Journal - Special Issue on AI in Design, Vol. 5, No. 1, (1992),
3-14
3. NATO SOFTWARE ENGINEERING CONFERENCE 1968, Garmisch, (1968)
4. Beck, K.: Extreme Programming Explained: Embrace Change, 224 pages ; Addison-Wesley
Pub Co; 1st edition, (1999)
On the Application of Biological Vision Concepts and
Mechanisms in very Demanding Industrial Visual
Processing
A. Quesada-Arencibia, J.C. Rodríguez-Rodríguez and R. Moreno-Díaz jr
Institute for Cybernetics (IUCTC), Universidad de Las Palmas de Gran Canaria

E35017 Las Palmas, SPAIN
aquesada@dis.ulpgc.es, jcarlos@ciber.ulpgc.es
rmorenoj@dis.ulpgc.es
Extended Abstract
The design of visual processing systems in very demanding industrial environments is

a technical field in which bioinspiration has not been explored as a developing tool.
The need of extremely quick, accurate and real time responses needed in industrial
applications usually is not seen as compatible with the “messy”, “slow” or
“inaccurate” methods and algorithms inspired in the information processing
mechanisms underlying neural activity in the visual pathway.
We are trying, thus, to explore the practical possibilities of interaction among
concepts from both worlds: the “real” vision system designed for a real time quality
control of a production line, and the “inspiration” taken from multi-channel biological
vision. In previous papers [1,2] a biologically plausible parallel system for visual
detection of form, movement, shape and size has been developed. The system,
working off-line and skipping real time restrictions, was tested for a variety of
situations, yielding very good results in estimating the mentioned visual
characteristics of moving objects. Furthermore, a second parallel-computing version
was design introducing the concept of parallel channel processing, e.g., the
discrimination of different visual characteristics by mean of multiprocessors and
multithread computing.
The architecture we present here, which includes certain concepts developed in the
previously explained results [3,4], is intended to work in the production line of a
beverage canning industry where cans with faulty imprinted use date and lot number
have to be immediately discharged from the line. The critical requirements being
accurate invariant from position text analysis in real time of nearly 30.000 cans per
minute, the system has to be reliable and rely on specific hardware. Thus, a high
speed acquisition camera, an efficient acquisition board, a strong multiprocessing
system and a considerable bandwidth for main memory load are the basic
requirements. Since the response of the system, e.g. whether a can is validated or not,
has to be immediate, it makes the actual visual processing time very small, sensibly
shorter than the 400 millisecond that biological systems need for a “good look”, that
is, for the effective processing of a visual frame in order to be able to recognize
specific objects or events [5].
In the following scheme, a biologically inspired architecture for this application is
presented. The basic ideas of multichanneling in the visual tract are present. On the
input image, a multiprocess task is first triggered to extract the area of interest where
first text is to be located. Thus, a second multichannel analysis analysis the possible
singularities in the text. The final validation consists of determining the coherence and
plausibility of text syntactically and semantically. All these processes are independent
and are separately operated. Thus, the labels {1}, {2}, {3} and {4} denote different
stages within the same visual tract.
C
A Acquisition Low Level Channel 1
M
Illumination 1
System
Text Presence Test
Low Level Channel n
Global and Coarse

Characteristics /Channels
Medium Level Channel 1
ENE 05 08:25
L4185/01
2
Medium Level Channel n Text Irregularities Test
Local and Coarse

High Level Channel 1

ENE 05 08:25
E$E0? 08:25
L4185/01
L4#85/0? 3
High Level Channel n Legibility Test
Very Local and fine

E$E0? 08:25 4
L4#85/0? Syntax Test
Code Validation
Semantic Test
Decision
Fig. 1. Multi-level and multi-channel architecture for validation
References
1 Alemán-Flores, M., Leibovic, K.N., Moreno Díaz jr, R.: A computacional Model for Visual
Size, Location and Movement, Springer Lectura Notes in Computer Science, Vol 1333.
Springer-Verlag. Berlin Heidelberg New York (1997) 406-419
2 Quesada-Arencibia, A., Moreno-Díaz jr, R., Alemán-Flores, M., Leibovic, K.N: Two
Parallel Channel CAST Vision System for Motion Analysis. Springer Lecture Notes in
Computer Science, Vol. 2178. Springer-Verlag. Heidelberg New York (2001) 316-327
3 Quesada-Arencibia, A.: Un Sistema Bioinspirado de Análisis y Seguimiento Visual de
Movimiento. Doctoral Dissertation. PhD Thesys. Universidad de Las Palmas de Gran
Canaria (2001)
4 J.C. Rodríguez Rodríguez, A.Quesada-Arencibia, R.Moreno-Díaz jr, and K.N. Leibovic: On
Parallel Channel Modelling of Retinal Processes Vol 2809 Springer-Verlag. Berlin
Heidelberg New York (2003) 471-481
5 Leibovic, K.N., Science of Vision, Springer Verlag, New York, 1990.
Theory-Building with System Dynamics: Principles and
Practices
Markus Schwaninger1, Thomas K. Hamann2

1
Professor of Management at the University of St. Gallen, Switzerland.
2
Senior Associate, McKinsey & Company, Inc., Munich, Germany
1 Introduction: Theory-building in Perspective
Theory-building, in principle, is more than an exercise in the ivory tower. It is an

activity fundamental to survival - for societies, organizations and individuals.
Constructing a model, in the sense used here, is building and mathematically
formalizing a theory in order to orientate action. It is a device for coping with the
complex. As complexities in our time tend to be high and often growing, the quest for
better theories is a necessity for both academics and practitioners.
Essentially, two modes of scientific inquiry can be distinguished - deductive
(concept-driven) and inductive (data-driven) theory-building. Typically, the former
leads to broad and logically compelling statements, which however lack empirical
corroboration, whereas the latter results in empirically well-founded insights that are
however restricted to narrow scopes. In order to overcome these limitations,
researchers have tried to combine the two modes of theory-building. This is often
found difficult or biased in one direction or the other. Hence, there is a need for
rigorous theory-building approaches, which balance out conceptual stringency with
empirical soundness.
In this direction, several attempts have been made, of which only two shall be
mentioned here. One is the approach of starting with an elaboration of - mostly broad
- theoretical concepts, which thereafter are submitted to empirical testing. The dyad of
the Academy of Management Review (conceptual orientation) and the Academy of
Management Journal (empirical orientation) is a case in point. The second case in
point is the combination of data gathering and concept formation in an iterative
process as practiced in the Grounded Theory approach founded by Glaser & Strauss
(1967).
In this contribution we demonstrate that the methodology of System Dynamics
offers a particularly powerful process and technique for effective theory-building in
order to improve decision-making in the context of organizations and society. This
methodology is designed to achieve both conceptual and empirical rigor.
System Dynamics (SD) is a discipline for the modeling, simulation and control of
complex dynamic systems, founded by Jay W. Forrester (1961, 1968). A main feature
of the SD modeling approach is that the issue modeled is represented by closed
feedback loops made up of essentially two kinds of variables - stocks and flows -
supplemented with parameters and auxiliary variables. The representation in form of
multiple closed loops as well as the consideration of delays, enable realistic modeling,
which brings the endogenous dynamics generated by the system itself to the fore.
Moreover, counterintuitive system behaviors (Forrester 1974) generated in the
simulations can lead to important insights for model users. Given these properties, SD
helps to avoid the so-called policy resistance -”the tendency for interventions to be
defeated by the response of the system to the intervention itself.” (Sterman 2001,
p. 8).
2 Theory-building with System Dynamics
The methodology of SD is centered around a process which combines modeling and

simulation iteratively, thus leading to a continuous improvement of model quality and
insights into the domain or issue modeled. Other authors have emphasized the role of
modeling as a vehicle for learning, in particular group learning, e.g., Lane (1994) and
Vennix (1996).
We take a new view by conceiving modeling and simulation as a powerful
approach to theory-building. An ideal-typical scheme of that process is depicted in
figure 1. Even though this is a general scheme, the process represented therein is
essentially a theory-building process with the sequence of formulating a proposition,
then testing it, then expanding or refining the proposition, proceeding with further
tests, etc.
The starting point is a framing of the issue at hand, including a coarse definition of
its scope and of the purpose of the model to be developed. The ensuing collection of
empirical data then, via a first view on reference patterns, supports the clarification of
the goals and the formulation of the research questions to be answered. From there, a
dynamic hypothesis can be formed, which explains the reference behavior pattern.
Besides empirical data, this dynamic hypothesis is also based on theoretical concepts
and constructs, which result from previous research efforts. The core of the theory-
building process then is about elaborating a theory by drawing on that dynamic
hypothesis, as well as testing, corroborating or refuting it. Model quality is enhanced
and explanation is deepened along this iterative process.
In the following, the theory-building along the various stages of this process will
be illustrated by instancing the generation of a holistic and consistent theory about the
development of individuals’ musical tastes, especially their preference for classical
music.
3 Application: The Case of Cultural Dynamics
It is crucial for the long-term success of all kinds of enterprises, including orchestras,
to discover and take opportunities as early as possible as well as to detect and avoid
potential threats before they become uncontrollable. Therefore, the anticipative
capturing of future realities by being able to interpret so-called weak signals regarding
external developments is important and might be critical for the long-term survival of
an organization (Ansoff 1975). Regarding the orchestra sector such a weak signal lies
in the fact that the German and Swiss classical music audience has a disproportionate
number of old people when compared to the population as a whole (significance level:
p = 0.001). The implications for the future size of the classical music audience and the
related demand for classical music life performances remained unclear, since there
was no theory about the formation of an individual’s affinity with classical music. In
particular, there was disagreement as to whether the fundamental musical taste is
predominantly subject to an aging/life cycle effect or a cohort effect due to the lack of
a theory. If people mature into listening and attendance practice as they become older
or by entering a new stage in the course of life (aging/life cycle effect), the classical
music audience will increase in the next decade. But if the younger adults with
decreased participation will not change their musical preferences over time (cohort
effect), the classical music audience will gradually die out. In order to fill this
research gap, one of the authors developed a theory about how the individuals’
musical preferences for classical music come into existence, by drawing on the SD
methodology (Hamann 2005).
Fig. 1: Ideal-typical scheme of an SD-based theory-building process
As a reference behavior pattern, the dynamic analysis of relevant empirical data

revealed that the classical music audience is aging faster than the population as a
whole. Moreover, the data prove the existence of a cohort effect.
Against this background the following dynamic hypothesis was formulated: The
activities with musical relevance - repeatedly listening to certain types of music,
playing an instrument, and attending appreciation classes - during a socializing phase
in the adolescence of an individual determine the fundamental musical orientation in
later life, since these activities enhance the development of “listening competence”,
i.e., what Behne (1997, p. 151) calls “cognitive components (‘concentrated’ and
‘distancing’)” of listening.
Simulations clearly corroborated the dynamic hypothesis. In addition deeper
insights into the issues under study were gained, which enabled the elaboration of
well-founded recommendations for the management of orchestras (Hamann 2005).
This case shows at least two features of the SD methodology. First, compared to
solely verbally formulated theories, SD’s mathematical formalization of the theory
leads to a more precisely specified theoretical body. This is because additional
hypotheses can derived mathematically, detailed formal consistency checks are
carried out, and all aspects relevant to the problem are covered systematically, which
in turn results in a deeper understanding of the problem and the identification of blind
spots that have to be covered by future research efforts. Therefore, and second, the SD
methodology enhances the validity of a theory.
4 On Model Validation
In theory-building, the quality of the theoretical propositions developed should be the

crucial concern. This is what the call for "scientific rigor" is all about. We are taking
Karl Popper's (1959) logic of scientific discovery - essentially a concept of an
evolutionary progress of science - as a benchmark for the design of the theory-
building process. This implies that any proposition must be formulated in such a way
that it can founder if confronted with reality. In other words: it must be proposed in a
way that it can be falsified. The tenet of this falsificationism is that science advances
by bold propositions or guesses followed by unstinting criticism. Only hypotheses
capable of clashing with facts count as scientific.
All of this is very much in line with the philosophy of SD. Bold guesses first
crystallize in the model which is then submitted to numerous tests, including various
simulation tests. This modeling and simulation process is inseparable of the validation
process - the totality of all measures to ensure the quality of a model. Validation is
about increasing the fit of a model to the reality it represents. This is a process of
gradually building confidence into the model (Barlas 1996), by both improvements of
structure and equations as well as the calibration of parameters and functions.
Among the methodologies for the modeling of social systems, as far as we know,
none has validation standards as strict as SD. For instance, econometrics operates
essentially with statistical validation procedures. In SD the standard procedure for
model validation also involves statistical tests. In order to avoid that a model is
considered right for the wrong reasons, SD validation includes a whole set of
procedures which are obligatory for a model to achieve credibility (see: Forrester &
Senge 1980, Barlas 1996, Sterman 2000).
5 Conclusions
Theory-building is more than an exercise for academics. It is a crucial device for

practitioners in organizations, allowing them to test their assumptions and bring their
speculations down to earth in order to make better decisions. That is why theory-
building is a fundamental prerequisite for effective action.
We have expounded the SD methodology as an exemplar of an approach to theory-
building. This discipline for the modeling, simulation and control is in line with the
concepts of evolutionary theory-building as proposed by the theory of science. It must
be added, that it is also highly appropriate for practical use due to the intuitive
techniques it uses and the user-friendly software available. The main limitations of
SD, on the other hand, are twofold. First, SD is appropriate for the modeling of
dynamic issues, but not of static ones. Second, that quantitative modeling always
requires quantitative data. In absence of those, only qualitative SD - by means of
feedback diagrams, i.e., what is usually called "Systems Thinking" - are applicable.
We have reported an application of SD to the building of a theory of Cultural
Dynamics, from which substantial insights and recommendations for the management
of cultural institutions have been derived. Other cases in point are already published,
e.g., Ulli-Beer (2004).
The potential fertility of SD as a methodology for theory-building is exceedingly
high.
Ontology Integration for Statistical Information
Wilfried Grossmann, Markus Moschner
Faculty of Computer Science, University of Vienna

WG Data Analysis and Computing
Universitaetsstrasse 5 A 1010 Vienna, Austria
{wilfried.grossmann, markus.moschner}@univie.ac.at
Extended Abstract
Statistical information plays a central role in many business and economic decisions.
The term information means here that we have to consider the data itself as well as
descriptions of the data, so called metadata, which are necessary for obtaining the
information. Hence, it is not surprising that metadata play a key role in statistical
information systems for a long time. The earliest reference is Sundgren [11], who
introduced the concept of metadata in statistics for a so called infological approach
towards statistical information. This approach has been developed further in many
ways by a number of researchers as well as statistical agencies and has lead to a
number of metadata standards. Due to the fact that a lot of statistical information is
contained in highly aggregated data, represented nowadays usually in data
warehouses, one line of development focussed on metadata standards for such type of
data (for example the SDDS-standard [9]). A second stream of development based on
ideas from documentation for data archives considered mainly the metadata
requirements for scientific researchers in social sciences and economics. These efforts
have resulted in different metadata standards, probably the best known example is the
DDI standard [1], which is a substantive expansion of the Dublin Core metadata [2].
A well known software tool based on these ideas is NESSTAR [7]. A further
development concentrated on proper metadata representation for value domains of the
attributes of interest and lead to the so called “Neuchatel Terminology” [8] for
classifications.
Due to the different starting points of these approaches it is rather cumbersome to
integrate data in cases where the definition of the data schemes is based on such
different documentation schemes. Following the developments of intelligent
information processing in recent years (see for example Sowa [10]) also in the field of
statistical information processing there were a number of efforts to develop the idea of
metadata further into the direction of ontology. In fact, statistical metadata practice
includes to a far extent information needed for ontology – engineering. Probably the
most important contribution in this direction was made by the METANET project (cf.
[4, 6]), a thematic network within the fifth EU-research framework. The approach
tries to fulfil the requirements of the ontology definition of Gruber [5] (“ontology” is
a specification of a conceptualization) by formalizing the statistical data paradigm,
taking into account the representational, operational and functional semantics of
statistical information.
Starting points are basic statistical objects like dataset, statistical population, or
statistical variable, which constitute the categories for the ontology. For all these
categories a unified description framework was developed, which is called the view
facet of statistical categories. The following four views were distinguished:
• The conceptual category view represents the subject-matter definition of any
category instance and builds the bridge to reality. Validity of the definition of the
subject matter concept gets usually restricted by temporal and geospatial
constraints.
• The statistical (methodological) category view describes the statistical properties
of the category instance by using a number of parameters, which have to be
defined in such way that specific properties of the different categories are taken
into account.
• The data management category view is geared towards machine supported
manipulation storage and retrieval of the category instance data.
• The administrative category view addresses management and book-keeping of the
structures.
Based on these view facets a representation scheme can be defined, which seems to
be sufficient for operational purposes. A first sketch of such a model was presented in
Denk et al. [3]. In this paper we present first and fundamental ideas for using this
framework for mapping different metadata standards.
References
1. DDI Data Documentation Initiative – A Project of the Social Science Community –

Codebook. http://www.icpsr.umich.edu/DDI/codebook/index.html.
2. DCMI Usage Board, DCMI Metadata Terms, 2003, Dublin Core Metadata Initiative
(DCMI), http://dublincore.org/documents/2003/03/04/dcmi-terms/.
3. Denk, M. Froeschl, K.A., Grossmann, W.: Statistical Composites: A Transformation
Bound Representation of Statistical Datasets. In: Kennedy, J. (ed.): Proc. 14th International
Conference on Scientific and Statistical Database Management, IEEE Los Alamitos,
California (2002), 217 – 226.
4. Froeschl, K.A., Grossmann, W., delVecchio, V.: The Concept of Statistical Metadata.
MetaNet (IST-1999-29093) Work Group 2, Deliverable 5 (2003).
5. Gruber, T. R.: A translation approach to portable ontologies. Knowledge Acquisition,
5(1993), 199-220.
6. MetaNet - http://www.epros.ed.ac.uk/metanet.
7. NESSTAR – Networked Social Science Tools and Resources. http://www.nesstar.org/.
8. Netterstroem, s: et al.: “Neuchatel Terminology – Classifications Data Base Object Types
and Their Attributes”, Version 2.0. Memo, September 2002.
9. SDDS – Special Data Dissemination Standard, Dissemination Standards Bulletin Board.
http://dsbb.imf.org/Applications/web/sddshome/#metadata.
10. Sowa, J. F. Ontology, Metadata, and Semiotics. In: Ganter, B., Mineau, G.W. (eds.):
Conceptual Structures: Logical, Linguistic, and Computational Issues. LNAI 1867,
Springer-Verlag Berlin (2000).
11. Sundgren, B.: An Infological Approach to Data Bases. Urval, nr. 7, National Central
Bureau of Statistics (Stockholm, Sweden, 1973).
Solving Influence Diagrams With Simulation
Methods
Arminda Moreno-Dı́az1 , Miguel Virto1 , Jacinto Martı́n2 , and David Rı́os Insua3
1
School of Computer Science. Madrid Technical University.
{amoreno, mvirto}@fi.upm.es
2
Dept. of Mathematics. University of Extremadura.
jrmartin@unex.es
3
Statistics and Decision Sciencies Group. Rey Juan Carlos University.
drios@escet.urjc.es
Abstract. Recent advances in graphical modeling tools and computa-

tional techniques have made simulation based methods a feasible alter-
native to solve complex real life problems. We center our attention on
decision making problems modeled by Influence Diagrams and describe
a simulation algorithm to find approximate solutions.
The importance of Bayesian Belief Networks ([6]) and Influence Diagrams

([3]) has been shown by their extensive use in modeling complex real life pro-
blems. This graphical modeling language for representing uncertainty, firmly
based on probability theory, is gaining acceptance in the computer science com-
munity. The increasing number of knowledge-based systems having these frame-
works as an integral part is a good example of their usefulness. In particular,
Influence Diagrams (ID), which are an enhaced version of belief networks, have
been extensively used in a variety of fields, from medicine ([5]) to finance ([4],
[1]) to military [7].
An ID structures a decision making scenario with its three basic elements:
decisions or actions to be made, uncertainties faced by the decision maker on
relevant variables of the system and a utility function that models preferences
over consequences resulting from the decisions made. As the optimal decision
path will depend on the way the random aspects play out, solving this sort of
problems will involve giving a decision rule which maximizes expected utility, for
every possible observable state, marginalizing over all other unobservable random
quantities. Methods for solving ID’s are numerous but as the complexity of the
diagrams increases, simulation based, rather than exact methods, are preferred.
This complexity may manifest, e.g., through large and continuous sets of actions,
intrincate relationships between model parameters and complex probability and
utility functions.
Within the Decision Analysis framework ([2]), our aim is to extend the work
in [9] and [10] to obtain approximate solutions of complex ID’s. The method
combines the augmented probability method ([8]), Markov Chain Monte Carlo
simulation techniques and a backward induction procedure to sequentially solve
the diagram. We start from a consistent diagram provided by an expert. An
initial simulation of possible realizations of the diagram will give clues of its
behaviour and tell us where to look for optimal policies. No preprocessing of the
diagram is needed as we work with sequences of nodes to perform simulations.
We provide new examples as well as a theoretical proof of the method.
References
1. Demirer, R., Charnes, J.M., Kellogg, D.: Influence Diagrams For Real Options Va-
luation. Working Paper No. 03-0805. SUIE School of Business (2003)
2. French, S., Rı́os Insua, D.: Statistical Decision Theory. Arnold (2000)
3. Howard, R.A., Matheson, J.E.: Influence Diagrams. In: R.A. Howard and J.E. Mat-
heson (eds.): Readings on the Principles and Applications of Decision Analysis. Vol.
2. Strategic Decision Group. Menlo Park (1984) 719–762
4. Lander, D., P.P.Shenoy.: Modeling and Valuing Real Options using Influence Dia-
grams. Working Paper No. 283. School of Business. University of Kansas. Lawrence,
KS (1999)
5. Nease, R.F., Dopuglas, K.O.: Use of Influence Diagrams to Structure Medical De-
cisions. Medical Decision Making 17 (1997) 263–276
6. Pearl, J.: Probabilistic Reasoning in Intelligent Systems. Morgan Kaufmann (1988)
7. Stafira, S. Jr., G. S. Parnell, J. T. Moore.: A methodology for evaluating military
systems in a counterproliferation role. Management Science 43 (1997) 1420-1430
8. Tanner, M.A., Wong, W.: The Calculation of Posterior Distributions by Data Aug-
mentation (with discussion). Journal of the American Statistical Association 82
(1987) 528550
9. Virto, M.A., Martı́n, J., Rı́os Insua, D., Moreno-Dı́az, A.: Approximate Solutions
of Complex Influence Diagrams through MCMC Methods. In: Proceedings of First
European Workshop on Probabilistic Graphical Models. Gámez, Salmerón (eds.)
(2002) 169–175
10. Virto, M.A., Martı́n, J., Rı́os Insua, D., Moreno-Dı́az, A.: A Method for Sequential
Optimization in Bayesian Analysis. Bayesian Statistics, 7. J.M. Bernardo, M.J. Ba-
yarri, J.O. Berger, A.P. Dawid, D. Heckerman, A.F.M Smith, M. West (eds.) Oxford
University Press (2003) 701–710
Decision support system using data warehousing
and on-line analytical processing technologies
Ryszard Klempous1 , Henryk Maciejewski1 , Jan Nikodem1 , Maciej Nikodem1 ,

Jerzy Rozenblit2 , Josef Scharinger3 , Carmen Paz Suarez Araujo4
1
WrocÃlaw University of Technology, Institute of Engineering Cybernetics
11/17 Janiszewskiego Street, 50-372 WrocÃlaw, Poland
2
University of Arizona, Electrical and Computer Engineering Department,
1230 E., Speedway Boulevard, Tuscon, USA
3
Johannes Kepler Universitaet Linz, Systemtheorie und Informationstechnik,
A-4040 Linz, Öesterreich
4
Universidad de Las Palmas de gran Canaria, Departamento de Informatı́ca y
Sistemas, Las Palmas, Las Palmas de Gran Canaria 35017, Spain
Extended abstract
Computer Assissted Assessment (CAA) of existing Information Technology (IT)

curricula will be discussed in this paper. Complete and permanent evaluation
of courses is being undertaken using special evaluation systems, development,
and deployment of software-based course evaluation systems. The assessment
tools that are being developed are based on data warehousing and data mining
technologies and will provide a means for integration of data from heterogeneous
environments as well as comprehensive reporting and multidimensional analysis
(OLAP) capabilities through an intuitive Web-based interface. The proposed de-
cision support system will allow for objective and comparable data-driven qual-
ity management of teaching in various countries and educational institutions
in Europe (Austria, Hungary, Poland, and Spain) and the USA (specifically at
the University of Arizona in Tucson). Data gathered by the system will make
it possible to analyze course quality based on student surveys and instructors’
analysis. A set of data-driven decision support tools will be developed for in-
tegrated quality management of educational processes. The tools will create a
platform for relating the quality of teaching processes from participating institu-
tions by taking into account comparative statistical analysis of teaching quality
data and disseminating results of analysis.
The system presented is being built using SAS System technology.
References
1. Dabrowski M., et al., Permanent Evaluation of Courses Using Specialized Evaluation

Systems, APLIMAT International Confernce Bratyslava, 2005 (to be appear)
2. Fayyad U.M., et al., Advances in Knowledge Discovery and Data Mining, MIT Press,
1996
3. Hastie T., et al., The Elements of Statistical Learning : Data Mining, Inference, and
Prediction, Springer Verlag, 2001
4. Klempous R., et al., The comparison of quality management of education process
based on data driven evaluation methods, ICSE ’2003, Coventry, 2003
5. Klempous R., et al., Quality management of education process based on data driven
evaluation methods, ITHET 2003, Marrakech, July 2003
6. http://www.SAS.com
Towards Automated Controlling of Human
Projectworking Based on Multiagent Systems
Manfred Mauerkirchner1 and Gerhard Hoefer1
Partner of University of Derby in Austria, A-4060 Leonding, Austria

m.mauerkirchner@eduhi.at
g.hoefer@htl-leonding.ac.at
Extended Abstract
Controlling a running project can be a most challenging task because at least

some parameters, which influence the project work flow, will certainly change
their values unpredictable. In addition, responsible shareholders should become
informed about potential risks which eventually could hinder achieving planned
results. To automate both - feedback concerning changing project resources and
permanent calculation of risks - we propose a flexible and scaleable model based
on multiagent methodology. At least there are two processes to observe: real
project flow and its realtime simulation based on our proposed model. Auto-
mated synchronization of both processes is the first goal we are trying to achieve.
We believe this is a suitable means to ensure that project resources can be allo-
cated, changed or deleted without corrupting the controlling system.
Because we focus mainly on persons as project resources (human resources)
we are forced to add several characteristic properties which are typical concern-
ing humans: e.g. specific skills/abilities to perform certain activities, heuristic
relationships in groups, etc. Finally, involving specific self learning components
should enable our system to store actual knowledge about both project and re-
sources in a transparent way, to approach our overall goal: automated controlling
of project working.The initial state is characterized by
– a directed graph of project activities (to define predecessor-successor rela-

tions between subtasks/activities) with estimated execution durations
– a set of human resources with different skill/cost profiles
– a matrix to describe heuristic relationships between persons
Starting the planning agent offline we calculate our first project plan, con-
taining a complete workflow of all activities and schedules of all involved persons
(Mauerkirchner, 2001). Subsequently both systems (real system and simulation
system) are ready to be started simultaneously based on our plan. In general, all
feedbacks stem from the real project system and enforce normally a recalcula-
tion of the actual project plan and, based on it a synchronisation of real system
and simulation system Feedbacks can be categorized in general feedbacks (e.g.
assignments or temporary absence of human resources) and personal feedbacks.
The latter ones come from persons using their human agent’s assessment inter-
face. Messages of isolated persons (groups with only one member) are sent to
the controlling agent at least (decision of project manager is necessary), while
feedbacks of other groups are received by the interference agent . It is possible to
create useful updates to the set of characteristic properties used by the central
planning agent. The bases for these updates are generated by extracting rele-
vant information from suggestions for improvement made by the human users
of the real system. These suggestions are transferred by the personal agent´s
assessment interface to the interference agent and stored as fuzzified statements
containing comments on problem relevant user skills and critical member inter-
actions in each user group. The collection of these statements is used by the
interference agent to generate an update of the above mentioned system prop-
erties in following distinct steps :
– create a population of sets that encode one statement for each characteristic
property, respectively.
– determine the fitness of each set by repeating offline the latest simulation of
the planning agent.
– choose sets that are able to improve the synchronisation of the actual project
flow in the real system and the simulated one.
– generate new populations of sets by performing crossover and mutations.
– repeat the last three steps until some improved steady state is reached.
The updated property values are checked by the controlling agent for plausi-
bility and then transferred to the planning agent for future simulations of project
flow. An alternative to this procedure might be the utilisation of Baysian Belief
Networks (Druzdel, 1996)(Pearl, 1988).
References
1. M.J.Druzdel, (1996) : Qualitative Verbal Explanations in Baysian Belief Networks.
AISB Quarterly1.96
2. J.Pearl, (1998): Probabilistic Reasoning in Intelligent Systems: Networks of Plausi-
ble Inference. Los Altos, CA: Morgan Kaufmann
3. M.Mauerkirchner, (2001): A General Planning Method for Allocation of Human Re-
source Groups, in Lecture Notes in Computer Science- Proceedings EURO-CAST’01
Las Palmas de Gran Canaria, Spain
On Recursive Functions and Well{founded
Relations in the Calculus of Constructions:
Extended Abstract
?
J. L. Freire, E. Freire, and A. Blanco

University of A Coru~
na, Department of Computer Science, LFCIA, Campus de
Elvi~
na, 15071 Coru~
na, Spain
ffreire,rike,blancog@dc.fi.udc.es
Abstract. As in Martin{L of's type theory general recursion is not avail-

able, a lot of work has been done to extend the theory to cope with this
powerful mechanism of program construction. Some examples are [2], [3],
[6]. The paper will describe the present situation of this problem and will
give some new methods and approaches.
We use the INRIA's implementation of the Calculus of Inductive Con-
structions Coq[5].
Key words: Well{founded Relations, Recursion, Type Theory, Inductive
Constructions
1 Introduction
Let a binary relation on a set A. The set Acc(A; ) of accessible elements of
is the set of elements a 2 A such that there is no innite descending sequence
n+1
:::a an :::a 2 a1 a:
The relation A A is well{founded if
A = Acc(A; ):
Formally we dene Acc(A; ) as a recursive predicate on A : an element x is
accessible if every y x is accessible.
Inductive Acc [A:Set;R:A->A->Prop]:A->Prop:=
Acc_intro:(x:A)((y:A)(R y x)->(Acc A R y))->(Acc A R x).
Definition well_founded := [A:Set;R:A->A->Prop](a:A)(Acc A R a)
In the paper detailed proofs of the justication of this representation are
given via informative versions of well{foundedness. Also, some connections with
intuitionistic and classical logic aspects of the models are analyzed.
? Government of Galicia, Spain, Project PGIDT02TIC00101CT and MCyT,
Spain, Project TIC 2002-02859, and Xunta de Galicia, Spain, Project
PGIDIT03PXIC10502PN
2 Results and Methods
Beginning from signicative examples we generalize to obtain some methods that
seem to be the base of a practical methodology to dene programs by general
recursion. This new methods are illustrated with some non trivial examples.
References
1. A. Abel, foetus Termination Checker for Simple Functional Programs. Techni-
cal Report. Theoretical Computer Science, Institute of Computer Science, Ludwigs{
Maximilians{ University (1999)
2. A. Abel, Termination Checking with Types. EDP Sciences (1999)
3. A. Balaa, Fonctions recursives generales dans le calcul des constructions. PhD.
Theese. Universite de Nice{Sophia Antipolis (2002)
4. A. Bove, Simple General Recursion in Type Theory. Technical Report. Chalmers
University of Technology, Goteborg (2000)
5. G. Doweck, A. Felty, H. Herbelin, G. Huet, C. Murty, C. Parent, C. Paulin{Mohring,
B. WernerThe Coq porrf assistant user's guide. Technical Report 134, INRIA. (1993)
6. B. Nordstrom, Terminating General Recursion. BIT, Vol. 28 (1988)

!
" #$%

&'%

( !
)
(
*

+ +,
++" *

-

.
/012

!
&& -

3

(

"

+

4

2

5
,

"

(

! 6&7 "
,
*

*

,
,

*

(

(

* 4
8
(
* "

9
,

, ,

89

(
)
(

*
)* ! 6&7

*
)* !

4 (

0 )* )*"

++" :
(
;)*"
;
)*"

:

;

;
)*"

(

;4 )*" (

;

(

.

)*" .

+
< . $

"
<<

)*
++"
2
, )*"

.
)*

++"

)*"
!
,

+"
+ ++"

(

/

,

.

/

& &
/
"

/

,

"

&

.

/ +

/

,

! !

"
#
$$ #%

,

&
= &
/67
/6 7>
/ ,
?

5 (

+ +,

@

,

.

,

++=

,

&

+ +,

&
@

,

.

/ 0$ !

,
6& &1 $& $0 017

& ' ( )
* +,

-

,

%
"

!. !/ & 0

%

1
02

# 3
# 45 % 6 3# $ 7

( $$89 9
# :9
#;

! << """&"=&">? "'$ 44+4 "$ +&5 @4"(1 44+4A
93# 7BC D E 9%, 3# D E 9%, 3#
, D E
89 D E 9 9
E 9FF;

G 93# 7B9C D E << 4&1 44+4 $&'$ @"1 @4"(1
H 89I D E 9 9 E 9IFF; << >' & )'4J44$$>'$ > & 44+4
/
<< @4 142 14 @4"(1 > &$$"( > 1"$ 44+4
89B9 C 9B9 C ::
9
B9 C D 9
B9 C ;
!
93# 7B9C D 93# 7B9IC F E

G
H << @4 244) '&2 > 14 44+4 @"1 1&$ +&5"++ @4"(1
/ 89
B9 C D 9
;

9%, 3# D 93# 7B9CE
9%, 3#
, D 9E
!

<< +K4' > 44+4$ " $4L44 D +&5 @4"(1
9
# D 9%, 3#E
G 9 D MME
H 9
E << &''&N$ $&' & @4"(1$ $&' & O
/ 9B9
C D 9B9
CE
! 9 D 9%, 3#
, 0 E
! 3# 8 9
; << @1"4 > & @4"(1$ >P4'4J
!
!! 893# 7B9C DD 9%, 3#;
!
! 9%, 3#E
! 9B9%, 3#C D 9 B9CE
!G
!H 9 E
!/

9E
<< "> 4J$ 14'4

$$ #%

Information Retrieval and Large Text
Structured Corpora ?
Fco. Mario Barcala1 , Miguel A. Molinero2 , and Eva Domı́nguez1

1
Centro Ramón Piñeiro, Ctra. Santiago-Noia km. 3, A Barcia,
15896 Santiago de Compostela, Spain
barcala@freeresearch.org, edomin@cirp.es
2
Dept. de Informática, Universidade de Vigo, Campus As Lagoas, s/n,
32004 Orense, Spain
molinero@uvigo.es
Extended Abstract
Conventional Information Retrieval Systems usually deal with plain text docu-
ments or with data files with a very elementary structure. Therefore, these kinds
of system are able to solve queries in a very efficient way, but only taking into
account whole documents, without distinguishing between different sections in
those documents. This behaviour makes it impossible to support queries that
should throw up a subset of documents which have some common characteristic
or perform searches based on document sections.
In contrast with this, text corpora commonly employed nowadays are usually
composed of a set of text files which in some way show a complex structure. So,
building a classical Information Retrieval System to work with this kind of data
will not benefit from this structure and results will not be improved.
In this article we test different technologies which can be used to build In-
formation Retrieval Systems to be applied on large structured corpora, namely,
technologies able to use the document structure to offer possibilities of filtering
the search. In addition, and as a result of our work, we can propose the technol-
ogy with the best flexibility-performance balance for use in building this kind of
system.
There are several examples of Information Retrieval Systems using struc-
tured corpora [1] [2], and many others about searching methods on large text
corpora [3], but there is no analysis or comparative study on technologies and
architectures that can be used to build them. An overview of different generic
technologies that could be applied to XML Information Retrieval Systems is
shown in XML Data Management, Native XML and XML-Enabled Database
Systems [4], but it is not about corpora.
In our analysis we suppose that the considered corpora are organized as a set
of different XML (Extensible Markup Language) [5] files. This assumption does
?
Partially supported by Ministerio de Educación y Ciencia (MEC) and FEDER
(TIN2004-07246-C02-01 and TIN2004-07246-C02-02), by MEC (HF2002-81),
and by Xunta de Galicia (PGIDIT02PXIB30501PR, PGIDIT02SIN01E and
PGIDIT03SIN30501PR)
not imply an important restriction since every text corpora can be transformed to
this format either automatically or manually, depending on the original structure
of the text.
Tamino 1 (a native XML indexer) [6], and Oracle2 [7] (a relational database
system which in its last versions integrates XML capabilities) are the main alter-
natives taken into account in our work. To perform our tests, we have taken into
account the following two main criteria: query flexibility and query performance.
Query performance evaluation is based on processing time measures using a
set of representative queries that covers a wide range of questions performed by
typical text corpora users. To evaluate query flexibility we have considered the
following aspects of the technologies under review:
– Statistical capabilities: The technology should allow the user to obtain

numerical values at different levels. For example, capabilities to count the
number of cases and documents which match each query.
– Additional information: It is also useful if the technology under review
is able to offer supplementary information with the returned results. For
example, returning sequences of terms which verify search criteria but also
showing the author of relevant document, publisher name, etc., i.e., addi-
tional data taken from corpora files structure.
– Match highlighting: Technologies should be able to identify the term
which caused the matching and highlight them among the retrieved results.
– Kinds of search methods: It is important to determine if a particular
technology suppports only exact matches or it allows other kinds of matching
(matching and indexing that is sensitive and insensitive to accents, boolean
search, complete or partial matching, etc.).
– Use of wildcards: Support of advanced features such as wildcards, char-
acter substitution, suppport for regular expresions, etc.
– Results browsing and navigation: The technology should also allow us
to go through the retrieved results, showing them page by page or even using
more sophisticated ways.
– Ordering: Ordering capabilities offered by the technology are also relevant.
Systems can allow single ordering criteria, several simultaneous orderings or
impose other limitations.
– Structural relationships: This aspect refers to query flexibility when struc-
tural restrictions are included in queries. This covers questions like capabil-
ities offered by the query language, and query complexity limitations.
The results obtained show that although XML native managers are a serious
candidate for the near future, they present certain deficiencies, which limit query
flexibility if we want to achieve a good performance. Only high performance
relational database systems like Oracle can satisfy some kind of requirements,
relevant to this kind of system.
1
Tamino is a Software AG product.
2
Oracle is a registered trademark of Oracle Corporation and/or its affiliates.
References
1. Marı́a Sol López Martı́nez CORGA (Corpus de Referencia del Gallego Actual) In
Proc. of Hizkuntza-corpusak. Oraina eta feroa, pages 500–504, Borovets, Bulgaria,
Sept. 2003.
2. Mark Davies Un corpus anotado de 100.000.000 palabras del español histórico y
moderno. In Proc. of Sociedad Española para el Procesamiento del Lenguaje Natural
pages 21–27, Valladolid, Spain, 2002.
3. Mark Davies Relational n-gram databases as a basis for unlimited annotation on
large corpora. In Proceedings from the Workshop on Shallow Processing of Large
Corpora, Lancaster, England, pages 23–33, Lancaster, England, March 2003.
4. Akmal B. Chaudhri, Awais Rashid and Roberto Zicari XML Data Management,
Native XML and XML-Enabled Database Systems, Addison-Wesley, March 2003.
5. XML In http://www.w3c.org, 27/10/2004
6. Tamino In http://www.softwareag.com, 27/10/2004
7. Oracle In http://www.oracle.com, 27/10/2004
Meteorological Image Descriptors
J. L. Crespo, P. Bernardos, M. E. Zorrilla, E. Mora
Department of Applied Mathematics and Computer Sciences, University of Cantabria.

Avda. de los Castros s/n 39005 Santander, Spain
{crespoj, bernardp, zorrillm, morae}@unican.es
Abstract. The objective of this paper is to get a visual characterisation of time

evolution images, in particular, synoptic maps taken from Meteorology. Pre-
liminary tasks required before image processing are reviewed. Two different
types of numerical descriptors are extracted for image characterisation, those
called low level numerical descriptors, and the corresponding high level ones.
The latter will subsequently be used for prediction tasks, while the former will
be used for classification tasks. Three different relevant information sources in
the images are identified as their low level descriptors. These are defined by the
local density and orientation of the isobar lines, and the number of centres of
high (H) and low (L) pressure. As regards the high level descriptors, two main
features are taken into account. The different procedures carried out to extract
the previous descriptors for the images of interest are discussed.
1 Introduction
The main aim of this paper is to obtain a visual characterisation of time evolution im-
ages, in particular, synoptic maps taken from Meteorology. As far as we know, no
similar work has been found to date in the computer vision literature for classification
or prediction of this kind of image, besides those that treat the problem with compli-
cated numerical models in Meteorology or statistical mathematical methods.
The preprocessing phase of the synoptic maps that we are working with has al-
ready been explained in [1]. Common preliminary tasks to single out the image in-
formation of interest, such as useless background removal, information separation and
memory usage have been discussed there.
2 Visual Characterisation of Synoptic Maps
The next step is studied in this paper and consists in extracting numerical descriptors
of the images in order to obtain a visual characterisation of them [2].
Since we are dealing with meteorological images, we focus on features that are re-
lated to atmospherical properties. In this sense, two types of numerical descriptors are
extracted: the low level and high level ones. The latter will be subsequently used for
prediction tasks, while the former will be used for classification tasks.
The low level numerical descriptors are the density of the isobar lines, their average
orientation in local areas, and the location of high (H) and low (L) pressure points.
The density of the lines is actually a downscaled image, following the approach of
pyramid representations [3]. For this purpose, we consider the isobar image to be di-
vided into adequate small windows. On the one hand, the size of the windows must be
small enough to identify the corresponding isobar image. However, on the other hand,
the window must be as big as possible in order to minimize the memory used per iso-
bar image, which will be important for the subsequent analysis, given the high num-
ber of images to be processed and stored.
The detection of H and L letters is a typical OCR task, but since no other letters are
involved, we use an ‘ad-hoc’ algorithm, instead of generic OCR techniques [chapter 2
of 4]. Simple methods do not work, because H and L letters present in our images are
not equal. For this reason, an analysis of the discriminatory power of several features
has been carried out.
To verify if the aforementioned extracted variables are good low level numerical
descriptors, we develop the inverse procedure, using them as the input for rebuilding
the synoptic map, and compare it with the original one. Satisfactory descriptions are
obtained for adequate window width.
Regarding the high level numerical descriptors, two main features are represented
per image: the big regions of locally profound high (H) and low (L) pressure and
those of nearly constantly oriented circulation.
Acknowledgments
The authors are deeply grateful to the Spanish Interministerial Board for Science and
Technology (Comisión Interministerial de Ciencia y Tecnología CICYT) which sup-
ports Project TIC2002-01306, as a part of which this abstract has been produced.
They would also like to thank the Meteorological Centre in EUVE (European Virtual
Engineering) Technology Centre (Vitoria, Spain) for their collaboration.
References
1. Crespo, J. L., Bernardos, P., Zorrilla, M. E., Mora, E.: Preprocessing Phase in the PIETSI
Project (Prediction of Time Evolution Images Using Intelligent Systems). 9th International
Workshop on Computer Aided Systems Theory: Eurocast 2003. Eds. Roberto Moreno-Díaz
and Franz R. Pichler. Lecture Notes in Computer Science 2809, 651-660. Springer-Verlag,
Berlin Heidelberg New York (2003)
2. Sonka, M., Hlavac, V., Boyle, R.: Image processing, analysis and machine vision. Interna-
tional Thomson Computer, London, Madrid (1996)
3. Shalkoff, R. J.: Digital Image Processing and computer vision. Wiley, New York (1989)
4. Cole, R. A. (ed. in chief): Survey of the State of the Art in Human Language Technology.
Cambridge University Press (1996)
Matching with a Hierarchical Ontology
—Extended Abstract—
Yaacov Choueka1,2 , Nachum Dershowitz1,3 , and Liad Tal1,4

1
Celebros, Ltd., 15 Abba Eban Blvd., Herzliya Pituach 46725, Israel
{yaacov,nachum,liad}@celebros.com
2
Department of Computer Science, Bar-Ilan University, Ramat Gan 52900, Israel
3
School of Computer Science, Tel Aviv University, Tel Aviv 69978, Israel
4
School of Mathematical Sciences, Tel Aviv University, Tel Aviv 69978, Israel
I now see the hierarchical as both beautiful and necessary.

—Camille Paglia (1992)
1 Introduction
Celebros (www.celebros.com) markets a search engine, QwiserTM [4], geared to

electronic commerce. Each user query retrieves a ranked list of matching prod-
ucts and a dialogue is initiated to refine (“drill down”) the search results down
to a short list of quality matches. For example, a search for “black bags” might
return hundreds of products. Since there are many different types of bags, in ad-
dition to displaying a ranked list of black bags, Qwiser allows an online customer
to choose from various categories, such as handbags, briefcases, doctor bags, and
backpacks. Products in the store’s catalogue are classified (either on-the-fly or
in a pre-processing stage) to each of the possible categories and subcategories.
See Fig. 1.
We describe here how hierarchical ontologies may be used for classifying
products, as well as for answering queries.
2 Hierarchical Ontologies
For classifying and scoring product descriptions, as well as queries, we make use
of declarative hypernym/hyponym (is-a) and meronym/holonym (part-of ) hier-
archies, such as the sense hierarchies provided by WordNet [3]. We assume these
ontologies are organized as directed acyclic graphs. Each node—representing
a class of objects (commodities and commodity categories, in our case) or an
attribute—may be related by is-a and part-of edges to other nodes. For exam-
ple: a rucksack is-a kind of backpack, which is-a kind of bag (in the sense of
container), jet black is-a shade of the attribute black, and a clasp is a part-of a
handbag, as is a handle.
Fig. 1. Sample search dialogue.
3 Classification
Making use of a hyponym taxonomy has long been a tool in hierarchical classi-
fication. See, for example, [5, 2, 6].
There are several novel ways in which an ontology can help:
1. Consider rival classifications, B (e.g. bag) and C (container ), of some item,

be they derived by linguistic or statistical means.
(a) If C is a hypernym of B, then the more specific classification B is pre-
ferred.
(b) If neither is a hypernym of the other (shirt and cardigan, say), then their
closest common ancestor (their least upper bound in the subset lattice)
A (garment) is a better classification than either.
2. Similarly for the meronym relation:
(a) If several sibling meronyms are competing, this suggests their shared
holonym (or any of its hyponyms), especially when the latter is also a
contender.
(b) To decide between rival meronym and holonym (lens and camera), ad-
ditional criteria are called for (e.g. the likelihood that a camera lens is
sold alone).
3. Even weak indications of relevance to a hypernym/hyponym/meronym/holo-
nym can help in word sense disambiguation.
For example, consider a product labelled “Totes Hooded Zip Front Jacket”
(“Totes” is a brand name) and described as follows: “This spring parka features
a zip front, two front pockets, a hood with a drawstring and an inside contrasting
pattern. The jacket is ideal for those warm, rainy days and is both functional
and fashionable.” Linguistic analysis suggests it may be a jacket (appears in the
name as well as in the catalog description) or a parka (appears in the beginning
of the descriptive paragraph). Since parka is a hyponym of jacket, the former,
more specific, category is preferred.
Also, the mention of pocket, a meronym of garment, which is a (distant) hy-
pernym of jacket, supports both interpretations. Since pocket is not a commodity,
it is not itself a contender. Furthermore, the words “parka” and “pocket” both
reinforce the garment sense of “jacket.”
Ambiguous queries may be subjected to similar analysis.
4 Retrieval
It is standard to organize retrieved items according to a predetermined hier-

archy. See [1]. The choices illustrated in Fig. 1 are dynamically chosen from a
hypernym/hyponym hierarchy.
There are several additional ways in which some search engines take advan-
tage of hierarchical ontologies. These include:
1. Retrieve all hyponyms (e.g. rucksacks and etuis) of query terms (“bag”),
but not their hypernyms.
2. If the query (“orange leather briefcase”) is for a subcategory in which there
are too few elements, one may wish to suggest siblings (orange leather back-
packs; red leather briefcases) or cousins from the hypernym hierarchy.
3. In some contexts, a query for a category might also propose its meronym
classes (lens for “cameras”) as also of relevance.
If hypernym/hyponym and meronym/holonym links are annotated with

meaningful weights, these can be used to assign probabilities to matches and
improve the quality of the ranking of retrieved items. For example, links from
W (e.g. watch) to its hypernyms I (instrument) and J (jewelry) may be weighted
by the likelihood that an item classified as W is an I or a J, respectively; the
reverse, hyponym link, from J to W , may be qualified by the chance that some-
one who asked for J will choose an item from W . Similarly, one may label the
link from meronym C (clasps) to holonym H (handbags) by the likelihood that
a C is part of an H and, in the reverse direction, that an H has part C.
5 Results
Preliminary experiments with many thousands of product descriptions (in En-

glish and German) are encouraging. Using a hypernym/hyponym hierarchy (as
in Section 3, Case 1 above), for example, improved linguistic classification in
about 83% of the cases in which it made a difference, and caused harm in only
2% of the cases considered. Depending on the depth of the ontology and on
the classification method used, the fraction of products that are affected varies
significantly.
Acknowledgements
We thank Michael Flor for his large share in this effort and Yael Villa for her
comments and encouragement.
References
1. Hao Chen and Susan Dumais. Bringing order to the Web: Automatically catego-
rizing search results. In Proceedings of the SIGCHI Conference on Human Factors
in Computing Systems, pages 145–152, The Hague, The Netherlands, April 2000.
ACM Press.
2. Susan T. Dumais and Hao Chen. Hierarchical classification of Web content. In
Nicholas J. Belkin, Peter Ingwersen, and Mun-Kew Leong, editors, Proceedings of
SIGIR-00, 23rd ACM International Conference on Research and Development in
Information Retrieval, pages 256–263, Athens, GR, July 2000. ACM Press.
3. Christiane Fellbaum, editor. WordNet: An Electronic Lexical Database. MIT Press,
Cambridge, MA, 1998.
4. Tal Rubenczyk, Nachum Dershowitz, Yaacov Choueka, Michael Flor, Oren Hod,
and Assaf Roth. Search engine method and apparatus. Patent pending.
5. Sam Scott and Stan Matwin. Text classification using WordNet hypernyms. In Pro-
ceedings of the COLING/ACL Workshop on Usage of WordNet in Natural Language
Processing Systems, pages 45–52, Montréal, Canada, August 1998.
6. Aixin Sun, Ee-Peng Lim, and Wee-Keong Ng. Hierarchical text classification meth-
ods and their specification. In Alvin T. S. Chan, Stephen C. F. Chan, H. V. Leong,
Vincent, and T. Y. Ng, editors, Cooperative Internet Computing, pages 236–256.
Kluwer Academic, Boston, March 2003.
Business Intelligence Process Through Information
Extraction with Lixto
José Aldo Díaz Prado
Database & Artificial Intelligence, Technische Universität Wien

Favoritenstr. 9 1040, Vienna, Austria
jadiaz@dbai.tuwien.ac.at, jadiaz@itesm.mx
Extended Abstract
Today, we are living in a competitive business world, the corporations need to be

more competitive that their competitors, so they need to know more in detail about
what is happening with their competitors in the business arena through the
information available in the internet and another data bases. Also the corporations
want to know, which kind of strategies the competitors are implementing and in
which directions they are moving. Who are the new customers?, what are the new
adquisitions?, what are the new mergers?, and what are the new strategies?, in
general, what is happening with each of their competitors.
To do this work, we propose to monitor the operations and results of each of these
corporations,we can extract this information by each of the websites of each of the
competitors (mainly formatted in HTML) in the World Wide Web using techniques
for data extraction applying Lixto [1]. The competitive business intelligence decisions
(sustained by relevant information) should help the organization to make more
effective strategic actions. The decision have larger implications for the ongoing
competitiviness of the company, impact on subsequent decisions, and are more
difficult to reverse that decisions of a more operational or tactical nature.
Key Intelligence
Topic
Planning and
directions
Needs Information Storage

and Processing
Decision Makers
Data
Dissemination Collection
Analysis and
Other Reporting
Users
Fig. 1. Analysis, Strategy, and Competitive Intelligence with Information

Intelligence is necessary to reduce uncertaint and risk in decision making. Often ,
the opinion available to an organization will be depend upon how early problems are
identified (we need to take the right information at the right time in the right place).
Competitive Intelligence encompasses the potencial effects (i.e., threats and
opportunities) created by all external elements of the business environment that
impact on the current competitiveness and future competitive ability of the
organization. It is a systematic process or cycle from collecting and analyzing
information about competitors activities (Fig. 1).
The solution proposed is thus to use wrapper technologies developed with Lixto to
extract the most relevant information from the websites of the competitors
(information in HTML) and then translate it into XML wich can be easily queried or
further processed. This information will be processed by other applications or sytems
and it will be integrated to the internal information systems into the corporation for
their further use (Repository of Competitors).
Wrappers are specialized program routines that automatically extract data from
Internet websites and convert the information into a structured format. More
specifically, wrappers have three main funtions. Firstly, they must be able to
download HTML pages from a website. Secondly, search for, recognise and extract
specified data. Third, save this data in a suitably structured format to enable futher
manipulation [2] . The data can be imported into another application for additional
processing. Over 80% of the published information on the Web is based on databases
runing in the background. When compiling this data into HTML document the
structure of the underlying databases is completly lost. Wrapper try to reverse this
process by restoring te information to a structure format. With the right programs, it is
even possible to use the Web as a large database. By using several wrappers to extract
data from the various information sources of the Web, the retrived data can be made
available in an appropriately structures format [3]. As a rule, a special development
wrapper is required for each individual data source, because of the different and
unique structure of websites.
The Web is also extremely dynamic and continually evolving, wich results in
frequent changes in the structures of websites. Cosequently, it is often necessary to
constantly update or even completely rewrite existing wrappers, in order to maintaine
the desired data extraction capabilities [4].
In specifying the structure of a page on the Web, two things need to be clearly
identified.
1. Tokens of interest on a page. By tokens i mean words or phrases that indicate the
heading of a section. A heading indicate the begining of a new section; thus
identifying headings identifies the section on a page.
2. The nesting hierachy with sections. Once a page has beeen descomposed in to
various sections, we have to identify the nesting structure of the sections.
The structuring task can be done automatically or with minimal user interaction.
The key here is that a program analyses the HTML and other formatting information
in a sample page from the source and guesses the interesting tokens on that page. The
system also use the formatting information to guess the nesting structure of the page.
The heuristics used for identifying important tokens on a page and the algorithm used
to organize sections into a nested hierarchy are an important for this work.
Tokens identifying the begining of a section are often presented in bold font in
HTML. They may also written enterely in upper case words, or may end with a colon.
After the tokens identification, the next step is to obtain the nesting hierarchy of
sections on a page, i.e., what sections comprise the page at the top level, what sub-
sections comprise other sections in the page, etc. As with the tokenizing step, the
nesting hierarchy can be obtained in a semi-actomatic fashion for a large number of
pages.
Querying the data in the WebSource requieres generating an appropriate URL
from the document that containe the answers. The answers are not strictly typed
objects in some standard representation. Instead, the answers have to be extracted
from the semi-structured data stored in the HTML documents, and these documents
may contain data that is irrelevant to the query. The structure of this documents may
be volatile and this affects the extraction process. Damain knowledge about the
WebSource is also embedded in the HTML documents and must be extracted.
Consider the Cemex ( Cementos Mexicanos) WebSources that offer investor
related information. An example of querying the WebSource with a Investor Center
binding is displayed in the figure 2. Additional information about the recent
developments and the NYSE is also showed in this WebSource. This WebSource
support a number of query bindings, e.g., Country, Customers.
For each set of allowed blindings, the translated query (URL) is different and we
need to specify the particular URL constructor expression for each binding. For a
query with a Investors binding, the URL is http://www.cemex.com/ic/ic_lp.asp.
Fig. 2. An example html document
Finally, structure data corresponding to the output type(s) must be extracted from
the answer HTML document wich contains irrelevant data as well.
A simple extractor processes an HTML document and constructs objects
corresponding to some output type specified in the corresponding row of the Wrapper
Capability Table. The contents of the HTML document are represented as an ordered
tree , so that declarative queries may be written to describe the data that is to be
extracted from the document. These queries use qualified path expressions to identify
nodes based on their contents; to traverse the ordered tree; and to extract data from the
elements (nodes). Additional operators of this languages provide the control to
manipulates data and to construct the output objects.
A Simple Data Model for HTML Documents
The contents of an HTML document and its structure (placement of the HTML
elements) are represented as an ordered tree, with the document as root. The nodes of
the tree correspond to the HTML elements of their components, e.g., table, list, img,
table row, list element, etc. The arcs of the tree represent the placement of the HTML
elements of their components within the document. The preorder enumeration of the
ordered tree corresponds to the top-down placement of HTML elements and their
components in the document. In a similar way to WebOQL [5], we represent an
HTML document as an ordered tree. However in contrast to WebOQL hypertrees, in
our representation each node is associated with a type wich represents an HTML
element or its components. The arc relate an HTML element with the list of HTML
elements embedded in it or the parameters that modify it. Our representation of an
HTML document allows us to simplify the query language of our extractors.
Consider of an HTML document whose ordered tree representation . It is
comprised of an HTML list followed by two HTML tables, each of wich is a node.
The components of these elements are also a nodes, e.g., table rows, (tr), and table
data (td). One td element comprises of an img element, wich is also a node.
In our data model, each HTML document (HTML element or parameter) is
described by a node in the ordered tree with the following attributes:
• Name: This represent the type of HTML element or its component, e.g., table, tr,
td, img, src, etc.
• Ocurrence: This is an integer representing the placement of this object in the
HTML document. The placement is recorder w.r.t. the HTML element that is the
parent node. The intergert value indicates the placement of this node w.r.t. the
sibiling node of the same type. In the figure 3. 4 the second table object has an
occurrence of 2, indicating it is the second table, even through it is the third child
of the parent (root).
• Data: This represent the optional data associated with an HTML element. The data
is particular to the type of element, for e.g., for the HTML element img, one
component is the scr parameter whose data is usually a URL.
• Child: This is a relationship that relates the object (HTML element or component)
represented by the node to the list of objects embedded in it or to the parameters
that modify it. It represents the arcs between a node and its children in the ordered
three. For example, the second ocurrence of the HTML element table in figure 3,
has two children of type tr. The second tr has one child of type, wich has one child
of the type img. The parameter src is represented as a child of the node img.
HTML document
root
List 1
Table2 Table1
Li 1 Li 2 tr 1 tr 2 tr 1 tr 2
td 1 td 2 table td 1
tr 1 tr 2 Img 1
src 1
td 1 td 1
Fig. 3. Tree representation of an HTML document
With this kind of representation our data extraction model will take sense and we
can do the intelligence process to take the relevant information of each of the websites
in XML, with these information we will feed the databases and we will can to develop
a corporate repository related with the transactional sytems in the organization.
References
1 Lixto Software GmbH, http://www.lixto.com

2 Golgher, P., Laender, A. And Ribeiro-Neto, B.: An Example-Based Environment
for Wrapper Generation, in: Proceeding of the 2nd International Workshop on The
World Wide Web and Conceptual Modeling, pp. 152-164, Salt Lake City, Utha,
USA, 2000
3 Eikvil, L.: Information Extraction from Wordl Wide Web – A Survey, Report
No.945, ISBN 82-539-0429-0, July 1999
4 Adelberg, B. And Denny, M.:Building Robust Wrappers for Text Sources,
Technical Report 1999,
http://www.cs.northwestern.edu/~adelberg/papers/www.pdf (Aug.2002)
5 G. Arocena, A. Mendelzon, “WebOQL: Restructuring Documents, Databases, and
Webs.” Proc. International Conference on Data Engineering, 1998.
Reasoning about a Functional Program for computing
Gröbner Basis?
J. Santiago Jorge, Víctor M. Gulías, José L. Freire, and Juan J. Sánchez
LFCIA Lab.
Department of Computer Science,
Universidade da Coruña
Campus de Elviña, s/n, 15071 A Coruña (Spain)
{sjorge,gulias,freire,juanjo}@dc.fi.udc.es,
http://www.lfcia.org
Extended Abstract
Proving the correctness of a program, even a simple one, is a complicated task and at
the same time it is one of the most important activities for a software engineer.
Functional programming [6,1,11] has often been suggested as a suitable tool for
writing programs that can be formally analyzed, and whose correctness can be as-
sured [5,8,7]. That assertion is due to referential transparency, a powerful mathematical
property of the functional paradigm that assures that equational reasoning, i.e., f x is al-
ways f x, makes sense. As a matter of fact, the mathematical way of proving theorems
can be successfully applied to computer science.
On the other hand, theorem provers help us in proving the correctness of a program.
These tools provide control and assistance in the processes of specification and prove;
thus they prevent bugs that could be introduced in an hand-made proof. The logical
framework Coq [3] is an implementation of the Calculus of Inductive Constructions
(CIC) [2,10], performed at INRIA. This system is a theorem prover goal-directed and
tactic-driven in the style of LCF [4].
The aim of this work is to contribute to the construction of a methodology to produce
certified software, i.e., we intend to find the way to strengthen two different notions:
First, there is a need of formal verification of the correctness of algorithms which goes
together with their construction; and second, programs are mathematical objects which
can be handled using logico-mathematical tools. These methodological techniques will
be exemplified by means of a program which calculates the Gröbner base of a set of
multivariate polynomials. We present the way in which one can formally verify several
aspects of the application implemented following the functional paradigm.
In this work, both the certification of the program and its efficiency will be taken
into consideration.
We present the proofs of properties in an informal (but exhaustive) style, with man-
ual proofs, where each step is justified by means of mathematical reasoning; and also,
the formalization of these proofs in the Coq proof assistant.
? This work has been partially supported by MCyT TIC 2002-02859 and Xunta de Galicia
PGIDIT03PXIC10502PN
First, we create the bases of the mechanization of an algebraic theory for the rings
of multivariate polynomials using dependent types.
Afterward we develop a theory about well-founded relations. A well-founded poly-
nomial ordering is formalized in the proof assistant Coq. Our aim here, is to establish
the accurate bases for proving the termination of certain reduction relations of polyno-
mials.
To conclude, an application in the field of symbolic computation is both developed
and verified: the calculation of the Gröbner basis for a set of multivariate polynomials.
We face a challenge: Reasoning about computer algebra programs requires the formal-
ization of an important amount of non-trivial mathematics.
We intend not only the formal certification of certain aspects of the program written
in the functional language Caml [9] but also the performance of efficient computing.
References
1. R. Bird and P. Wadler. Introduction to Functional Programming. Prentice Hall, 1987.
2. T. Coquand and G. Huet. The Calculus of Constructions. Information and Computation,
1988.
3. G. Dowek, A. Felty, H. Herbelin, G. Huet, C. Paulin-Mohring, and B. Werner. The COQ
Proof assistant user’s guide. Technical report, INRIA, 1991.
4. M. J. Gordon, A. J. Milner, and C. P. Wadsworth. Edinburgh LCF: A Mechanised Logic of
Computation. Springer-Verlag, 1979.
5. V. M. Gulías, J. S. Jorge, , J. J. Quintela, and J. L. Freire. From Sequential to Distributed
Formal Proofs. In F. Pichler and R. Moreno-Díaz, editors, Computer Aided Systems The-
ory (EUROCAST’97), volume 1333 of Lecture Notes in Computer Science. Springer-Verlag,
1997.
6. P. Hudak. Conception, Evolution, and Application of Functional Programming Languages.
ACM Computing Surveys, 1989.
7. J. S. Jorge. Estudio de la verificación de propiedades de programas funcionales: de las prue-
bas manuales al uso de asistentes de pruebas. PhD thesis, Departamento de Computación,
Facultade de Informáticas, Universidade Da Coruña, oct 2004.
8. J. S. Jorge and J. L. Freire. Prove and then Execute. In 7o Encuentro de Álgebra Computa-
cional y Aplicaciones (EACA-01), 2001.
9. X. Leroy. The Objective Caml System: Documentation and User’s Manual, 2001. V3.04.
10. C. Paulin-Mohring. Inductive definitions in the system Coq: Rules and properties. In
M. Bezem and J. F. Groote, editors, Typed Lambda Calculi and Applications (TLCA’93),
volume 664 of Lecture Notes in Computer Science. Springer-Verlag, 1993.
11. L. C. Paulson. ML for the Working Programmer. Cambridge University Press, 2nd edition,
1996.
CheapTB: A Low Cost of Operation Distributed
File System
Javier Parı́s, Victor M. Gulı́as, and Carlos Abalde

{paris,gulias,carlos}@dc.fi.udc.es
Laboratorio de Fundamentos de Computación e Inteligencia Artificial, MADS Group.

Universidade de A Coruña
There are several applications that have been developed in recent years which
need great amounts of storage space, often bigger than what a single worksta-
tion can provide. Examples of this applications are the storage of digital video,
scientific experiments data or weather forecasting.
The traditional way of solving this problem is the use of some costly storage
service, like a SAN. Business which sell those systems often include maintenance
so this is an adequate solution for companies and research centers without budget
constraints. However, this big costs make it difficult for small business or research
labs to use this sort of solutions as they cannot afford them. The objetive of
CheapTB is to offer a distributed filesystem whose focus is low cost of adquisition
and operation rather than high performance.
As the storage capacity of IDE and SCSI disk grows it is increasingly possible
to put several disks in one machine. A cluster of these machines can be used as a
storage service. The cost of a storage cluster like the one described is much lower
than that of a SAN solution, although performance will be lower. Still, there
are many applications which need high storage space that don’t have critical
requirements in throughput or latency.
Some software is needed to use all of the space offered by the disks as if it
was a single disk. This software must manage the storage of files in the cluster
distributing them between the different disks the machines have, in several steps:
– It must be accessible through some standard network filesystem protocol,

like NFS[2, 3] or SMB[4].
– It must translate these accesses in data retrievals from or data storage into
the different disks the system has.
– It must manage accesses to the disks in all the storage nodes.
As this software controls the way data is stored in the cluster, it can control
power usage by the system powering off disks and machines which are not in use
at a certain time. This makes the system cheaper to operate.
Although it is possible to shut down machines in the cluster which are not in
use, at least a control node must always remain on to answer to external requests.
This node will be the same that translates external requests into accesses to the
different nodes, as it must know where all the data is in order to know which
machines must be powered on to answer an external request.
The system has an structure in three layers. The first layer is the interface
with the user. This layer offers access to CheapTB through several common
protocols, like NFS or SMB, using protocol adapters. CheapTB is designed so
that it is easy to add new protocols as needed.
These protocol adapters work on top of a common API with the second layer,
which is the filesystem. The API is similar to POSIX[1], to make adaptation of
existing network filesystem servers easy.
The filesystem manages the storage of the data and keeps track of the direc-
tory structure. The data is stored in a virtual block device, which is the third
layer. This virtual block device hides the complexity of managing all the different
storage media and the distributed nature of the system from the upper layers.
The power management works by distributing the responsability between the
filesystem, which has all the information about files and directory structure, and
the virtual block device, which has all the information about devices and nodes.
References
1. IEEE, ”Institute of Electrical and Electronics Engineers”, Standard 1003.1, 2003
Edition http://www.unix.org/single unix specification/
2. Sun Microsystems, Inc., NFS: Network File System Protocol Specification
http://www.ietf.org/rfc/rfc1094.txt
3. B. Callaghan, B. Pawlowski, P. Staubach and Sun Microsystems, Inc., NFS Version
3 Protocol Specification, http://www.ietf.org/rfc/rfc1813.txt
4. Christopher R. Hertel, Implementing CIFS, http://www.ubiqx.org/cifs/
Spelling Correction on Technical Documents?
M. Vilares1 , J. Otero1 , and J. Graña2

1
Department of Computer Science, University of Vigo
Campus As Lagoas s/n, 32004 Ourense, Spain
{vilares,jop}@uvigo.es
2
Department of Computer Science, University of A Coruña
Campus de Elviña s/n, 15071 A Coruña, Spain
grana@udc.es
1 Introduction
Spelling correction is a critical task in a variety of professional applications

related to the elaboration of technical documents, for which efficiency, safety
and maintenance are properties that cannot be neglected.
In contrast to fully automatic correction [3], most commercial systems
assist users by offering a set of candidate corrections that are close to the
misspelled word. So, any technique allowing the repair process to reduce the
number of candidates considered for correction implies improved efficiency and
maintenance, which should not have any negative effect on safety. In order to
ensure this, we focus on limiting the size of the repair region in the word, in
contrast to previous proposals considering a global region. As a result, we work
towards obtaining a reduction in complexity and enhanced performance, without
loss of repair quality.
2 Asymptotic behavior
We introduce a regional repair technique to deal with spelling correction on

technical texts, a field where isolated-word strategies can be preferable to
context-sensitive ones and, as a result, morphological aspects strongly impact
performance. We choose to work with the Itu corpus1 , the main collection of
texts about telecommunications. Here, we have considered a sub-corpus of 17.423
words that have been used as a part of the support for the Crater project [1].
The recognizer is a finite automaton (fa) containing 11.193 states connected
by 23.278 transitions built from Galena [2]. From this sub-corpus, we have
selected a representative sample with the same distribution in terms of lengths
of the words dealt with. For each length-category, errors are randomly generated
in number and position.
?
Research supported by the Spanish Government under projects TIN2004-07246-C03-
01, TIN2004-07246-C03-02 and HP2002-0081, and the Autonomous Government of
Galicia under projects PGIDIT03SIN30501PR and PGIDIT02SIN01E.
1
For International Telecommunications Union CCITT Handbook.
2.1 The running language
Our running language is Spanish, with a great variety of morphological processes,

making it adequate for our description. The most outstanding features are to
be found in verbs, with a highly complex conjugation paradigm, including nine
simple tenses and nine compound tenses, all of which have six different persons.
If we add the present imperative with two forms, the infinitive, the compound
infinitive, the gerund, the compound gerund, and the participle with four forms,
then 118 inflected forms are possible for each verb. In addition, irregularities are
present in both stems and endings. So, very common verbs, such as hacer (to
do), have up to seven different stems: hac-er, hag-o, hic-e, har-é, hiz-o, haz,
hech-o. Approximately 30% of Spanish verbs are irregular, and can be grouped
around 38 different models. Verbs also include enclitic pronouns producing
changes in the stem due to the presence of accents: da (give), dame (give me),
dámelo (give it to me). We have considered forms with up to three enclitic
pronouns, like tráetemelo (bring it for you and me). There exist some highly
irregular verbs that cannot be classified in any irregular model, such as ir (to
go) or ser (to be); and others include gaps in which some forms are missing or
simply not used. For instance, meteorological verbs such as nevar (to snow) are
conjugated only in third person singular. Finally, verbs can present duplicate
past participles, like impreso and imprimido (printed).
1e+06 1e+07
Our proposal Our proposal
Savary’s proposal Savary’s proposal
Number of computed items
1e+06
100000
100000
10000
10000
1000
1000
100
100
10 10
0 2 4 6 8 10 12 14 16 16 14 12 10 8 6 4 2 0
Length of the current suffix Position of the first point of error
Fig. 1. Number of items generated in error mode. Logarithmic scale.
This complexity extends to gender inflection, with words considering only

one gender, such as hombre (man) and mujer (woman), and words with the
same form for both genders, such as azul (blue). In relation to words with
separate forms for masculine and feminine, we have a lot of models: autor,
autora (author); jefe, jefa (boss); poeta, poetisa (poet); rey, reina (king)
or actor, actriz (actor). We have considered 20 variation groups for gender.
We can also refer to number inflection, with words presenting only the singular
form, as estrés (stress), and others where only the plural form is correct, as
matemáticas (mathematics). The construction of different forms does not involve
as many variants as in the case of gender, but we can also consider a certain
number of models: rojo, rojos (red); luz, luces (light); lord, lores (lord) or
frac, fraques (dress coat). We have considered 10 variation groups for number.
2.2 The computational cost

We compare our proposal with the Savary’s global approach [4], to the best of
our knowledge the most efficient method of error-tolerant look-up in finite-state
dictionaries. We consider the set of calculations associated to a transition in the
fa, that we call item, as unity to measure the computational effort. Finally, the
precision will reflect when the correction expected by the user is provided.
Some preliminary results are compiled in Fig. 1. They illustrate our
contribution from two viewpoints: First, our proposal shows a linear-like
behavior, in contrast to the Savary’s one, which seems to be of an exponential
nature. Second, the number of items is significantly reduced when we apply our
regional criterion. These tests provided a precision of 77% (resp. 81%) for the
regional (resp. global) approach. The integration of linguistic information should
reduce this gap, which stands at less than 4%, or even eliminate it. In effect, our
regional approach currently only takes into account morphological information,
which has an impact on precision, while a global technique does not exclude any
repair alternative.
2.3 The performance

However, statistics on computational cost only provide a partial view of the
repair process that must also take into account data related to the performance
from both the user’s and the system’s viewpoint. In order to get this, we have
introduced the following two measures, for a given word, w, containing an error:
useful items proposed corrections
performance(w) = recall(w) =
total items total corrections
that we complement with a global measure on the precision of the error repair
approach in each case, that is, the rate reflecting when the algorithm provides
the correction attended by the user. We use the term useful items to refer to the
number of generated items that finally contribute to obtain a repair, and total
items to refer to the number of these structures generated during the process.
We denote by proposed corrections the number of corrections provided by the
algorithm, and by total corrections the number of possible ones, absolutely.
These results are shown in Fig. 2, illustrating some interesting aspects in
relation with the asymptotic behavior we want to put into evidence in the
regional approach. So, considering the running example, the performance in our
case is not only better than Savary’s; but the existing difference between them
increases with the location of the first point of error.
A similar behavior can be observed with respect to the recall relation. Here,
Savary’s algorithm shows a constant graph since the approach applied is global
and, as consequence, the set of corrections provided is always the entire one
for a fixed error counter. In our proposal, the results prove that the recall is
smaller than for Savary’s, which illustrates the gain in computational efficiency
in opposite to the global method. Related to the convergence between regional
and global approaches, we must again search around points of detection close to
the beginning of the word, which often also implies repair regions be equivalent
to the total one and repairs starting around of q0 , such as is illustrated in Fig. 2.
1
Our proposal Our proposal
Savary’s proposal 1,4 Savary’s proposal

0,8
1,2
0,6
1
0,4
0,8
0,2
0,6
0
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 18
Position of the first point of error Position of the first point of error
Fig. 2. Performance and recall results.
Finally, the regional (resp. the global) approach provided as correction the
word from which the error was randomly included in a 77% (resp. 81%) of
the cases. Although this could be interpreted as a justification to use global
methods, it is necessary to remember that we are now only taking into account
morphological information, which has an impact in the precision for a regional
approach, but not for a global one that always provide all the repair alternatives
without exclusion. So, the consideration of the precision concept represents, in
the exclusive morphological context considered, a clear disadvantage for our
proposal since it bases its efficiency in the limitation of the search space. We
attend that the integration of linguistic information from both, syntactic and
semantic viewpoints will reduce significantly this gap, less than 4%, around the
precision; or even will eliminate it.
References
1. Thorsten Brants. Some experiments with the CRATER corpus. Technical report,
Universität des Saarlandes, Saarbrücken, 1995.
2. J. Graña, F.M. Barcala, and M.A. Alonso. Compilation methods of minimal acyclic
automata for large dictionaries. Lecture Notes in Computer Science, 2494:135–148,
2002.
3. K. Kukich. Techniques for automatically correcting words in text. ACM Computing
Surveys, 24(4):377–439, December 1992.
4. A. Savary. Typographical nearest-neighbor search in a finite-state lexicon and its
application to spelling correction. Lecture Notes in Computer Science, 2494:251–
260, 2001.
Verification of Language Based Fault-Tolerance
Clara Benac Earle1 and Lars-Åke Fredlund2

1
Computing Laboratory, University of Kent, England
2
Swedish Institute of Computer Science, Stockholm, Sweden
Extended Abstract
As software based critical systems are now becoming widely deployed, it is a

crucial development task to analyse whether such systems will survive the in-
evitable faults (in hardware, in network links, in software components) that
occur. However, this remains very difficult. It is hard to develop realistic failure
models for simulation and testing (often the environment characteristics are not
fully known), and to test and simulate for all possible faults would be very time
consuming. Consequently, here is an area where there is a need for formal verifi-
cation. But that too is hard. Most earlier work on verifying fault-tolerance target
a single application only, are ad-hoc, and do not provide a reusable verification
method. In this paper, instead, we introduce a verification method based on
model checking, that, since it addresses programs developed using higher-level
design patterns which address fault-tolerance in a structured way, can be reused
for a large set of such applications.
The Erlang programming language [1] is unusual in several respects. First, it
is one of a few functional languages, with extensions for concurrency and distri-
bution, that is used in industry. At first used only inside the Ericson corporation,
the language has by now spread widely outside Ericson, and a number of im-
pressive, complex, applications with high-availability requirements [2] have been
developed using the language (see http://www.erlang.org for more informa-
tion). Secondly the programming language itself includes constructs for handling
fault-tolerance, and it is fair to say that no small part of the popularity of Erlang
is due to the focus on fault-tolerance.
In brief the Erlang programming model for distribution is based on processes
communicating using message-passing; each process executes a function, which
has no side effects except sending and receiving messages from and to other
processes. On top of the core language a library of design patterns (OTP) for
implementing distributed applications has been added [3]. The library defines,
for instance, a pattern for programming server processes which provides generic
support for uniform client access, handling of fault-tolerance, and permits code
updates during runtime.
In the core language fault-tolerance is handled on two levels. First, as is
standard, the language has primitives for trapping and raising exceptions (error
conditions) inside a process (catch and throw). More interesting is the intro-
duction of inter-process fault-tolerance handling in the language. Two processes
can be bidirectionally linked, so that if one of the processes terminates abnor-
mally, the second process is notified. In case the second process takes no special
action, it will also terminate abnormally as a result of receiving the notification.
Thus, among linked processes, abnormal terminations will eventually kill off all
processes. A process can elect to receive the abnormal termination notification
in its message mailbox instead, and will then not be terminated.
On top of this low-level inter-process fault handling mechanism, the OTP li-
brary provides a higher-level design pattern, a supervisor structure, which orders
processes in a tree-like structure for error detection and recovery in a hierarchical
manner. If a lower-level process terminates abnormally, it is up to its immediate
supervisor (the process one level up in the supervisor tree structure) to take cor-
rective action, i.e., either to (i) restart the process, or (ii) restart all its children
processes (if they are interdependent), or (iii) decide that it cannot resolve the
problem and instead terminate abnormally itself, thus handing over handling of
the error condition to its parent supervisor one level up.
The Erlang systems we analyse for fault-tolerance are built from a set of
such generic server components, organised in a supervisor structure, and a set
of client processes accessing the servers.
We have adapted an existing verification toolset [4] to the task of verifying
correct behaviour under faults. The verification toolset has several components.
First Erlang systems are translated into µCRL specifications [5] (a process al-
gebra with data). Next the µCRL toolset3 generates the state space from the
algebraic specification, and finally the CADP toolset4 is used to check whether
the system satisfies correctness properties specified in a temporal logic.
To enable analysis of fault behaviour we introduce during the translation
phase to µCRL explicit failure points in the algebraic specification, in a system-
atic way, where the system processes may fail. The key observation of this paper
is that, due to the usage of higher-level design pattern that structure process
communication and fault recovery, the number of such failure points that needs
to be inserted can be relatively few, and can be inserted in an appplication inde-
pendent manner. In other words, the state spaces generated from a failure model
can be generated automatically, are relatively small, and are thus amenable to
model checking.
We demonstrate the approach in a case study where a server, built using
the generic server design pattern, implements a locking service for the client
processes accessing it [6]. The server necessarily contains code to handle the
situation where clients can fail; if it did not the server would quickly deadlock.
In the study we verify, using the automated translation and model checking
tool, systems composed of a server and a set of clients with regards to crucial
correctness properties such as deadlock freedom, mutual exclusion and liveness.
3
http://homepages.cwi.nl/ mcrl/mutool.html
4
http://www.inrialpes.fr/vasy/cadp/
References
1. Armstrong, J., Virding, R., Wikström, C., Williams, M.: Concurrent Programming
in Erlang (Second Edition). Prentice-Hall International (UK) Ltd. (1996)
2. Wiger, U.: Four-fold increase in productivity and quality – industrial-strength func-
tional programming in telecom-class products. In: Proceedings of the 2001 Workshop
on Formal Design of Safety Critical Embedded Systems (FEmSYS 2001). (2001)
3. Torstendahl, S.: Open telecom platform. Ericsson Review 1 (1997)
4. Arts, T., Earle, C.B., Penas, J.J.: Translating Erlang to mucrl. In: Proceedings
of the International Conference on Application of Concurrency to System Design
(ACSD2004), IEEE Computer Society Press (2004)
5. Groote, J.: The syntax and semantics of timed µCRL. Technical Report SEN-R9709,
CWI, Amsterdam (1997)
6. Arts, T., Earle, C.B., Derrick, J.: Development of a verified Erlang program for
resource locking. International Journal on Software Tools for Technology Transfer
(STTT) 5 (2004) 205–220
Applying Stacking and Corpus Transformation
to a Chunking Task
José A. Troyano, Vı́ctor J. Dı́az, Fernando Enrı́quez and Vicente Carrillo
Department of Languages and Computer Systems

University of Seville
Av. Reina Mercedes s/n 41012, Sevilla (Spain)
troyano@lsi.us.es
1 Extended abstract
There are many tasks in natural language processing that consist of associating
some kind of category to a group of words. Named Entity Extraction, Shallow
Parsing or Semantic Role Identification are three good examples. In this type
of tasks, we can identify two subtasks: one that finds the limits of the group of
words (chunk) and a second process that associates the correct tag to this group.
In this paper we present a series of experiments on a clear example of chunking:
the NER (Named Entity Recognition) task. We show how corpus transformation
and system combination techniques improve the performance in this task.
The NER task consists in the identification of the group of words that form a
named entity. IOB notation is usually employed to mark the entities in a corpus.
In this notation, the B tag denotes the beginning of a name, the I tag is assigned
to those words that are within (or at the end of) a name, and the O tag is
reserved for those words that do not belong to any named entity.
In the development of our experiments we have used the Spanish corpus dis-
tributed for the CoNLL’02 shared task [4], and the re-trainable tagger generator
TnT [1] based on Markov Models. In order to improve the performance of the
NER task we have increased the number of models. We have defined three trans-
formations that give us additional versions of the training corpus, and we have
trained TnT with them to obtain different taggers. The three transformations
improve the baseline given by TnT trained with the original corpus. Then we
have applied a stacking (machine learning) scheme to combine the results of the
models. The results of the experiments are presented in Table 1, we rely on the
well known recall, precision and Fβ=1 measures to compare the results.
Vocabulary reduction (Tnt-V): This transformation discards most of the in-

formation given by words in the corpus, emphasizing the most useful features for
the recognition. We employ a technique similar to that used in [3]. Each word
is replaced by a representative token, for example, starts cap for words that
start with capital letters, these word patterns are identified using a small set
of regular expressions. Not all words are replaced with its corresponding token,
trigger words remain as they appear in the original corpus.
Change of tag set (TnT-N): We replace the original BIO notation with a
more expressive one that includes information about the words that usually end
a named entity. We have added two new tags E and BE that are assigned, respec-
tively, to words that end a multi-word named entity and to single-word named
entities.
Addition of part-of-speech information (TnT-P): In this case we will make

use of external knowledge to add new information to the original corpus. Each
word will be replaced with a compound tag that integrates two pieces of infor-
mation, the result of the first transformation (vocabulary reduction) and the
POS tag of the word.
Figure 1 shows the result of the application of the transformations to a corpus

fragment
Word Tag Word Tag Word Tag Word Tag

La O La O La O La det O
Delegación B starts cap B Delegación B starts cap noun B
de I de I de I de prep I
la I la I la I la det I
Agencia I starts cap I Agencia I starts cap noun I
EFE I all cap I EFE E all cap noun I
en O en O en O en prep O
Extremadura B starts cap B Extremadura BE starts cap noun B
transmitirá O transmitirá O transmitirá O transmitirá verb O
hoy O lower O hoy O lower adv O
... ... ... ... ... ... ... ...
a) Original b) Vocabulary c) Change of d) Addition of POS
corpus. reduction. tag set. information.
Fig. 1. Result of applying the three transformations.
System Combination through Stacking (TnT-Stack): Stacking consists in ap-

plying machine learning techniques for combining the results of different models.
The final decision is taken according to a pattern of correct and wrong answers.
We need a training database to be able of learning the way in which every model
is right or wrong. Each register in the training database includes all the tags pro-
posed by the participant models for a given word and the actual tag. With this
kind of combination we reach a Fβ=1 measure of 88.37 improving more than 6
points the baseline. This performance is similar to state of the art recognizers,
with comparable results to those obtained by one of the best NER systems for
Spanish texts [2].
Table 1 shows the results of our experiments. The baseline NER-baseline is
defined by training TnT with the original version of the training corpus. The
table also shows the results obtained with the three transformations (experiments
TnT-V, TnT-N and TnT-P) and the the results of applying the stacking scheme
(TnT-Stack).
Precision Recall Fβ=1

NER-baseline 81.40% 82.28% 81.84
TnT-V 81.76% 85.59% 83.63
TnT-N 85.04% 84.15% 84.59
TnT-P 81.51% 84.79% 83.12
TnT-Stack 88.68% 88.05% 88.37
Table 1. Results of NER experiments.
References
1. Brants, T.: TnT. A statistical part-of-speech tagger. In Proc. of the 6th Applied
NLP Conference (ANLP00). USA (2000) 224–231
2. Carreras, X., L. Màrquez y L. Padró: Named Entity Extraction using AdaBoost. In
CoNLL02 Computational Natural Language Learning. Taiwan (2002) 167–170
3. Rössler, M.: Using Markov Models for Named Entity recognition in German news-
papers. In Proc. of the Workshop on Machine Learning Approaches in Computational
Linguistics. Italy (2002) 29–37
4. Tjong, E.F.: Introduction to the CoNLL’02 Shared Task: Language-Independent
Named Entity Recognition. In Proc. of CoNLL-2002. Taiwan (2002) 155–158
A Formally Verified Proof (in PVS) of the
Strong Completeness Theorem of Propositional
SLD–resolution?
J.A. Alonso, M.J. Hidalgo, F.J. Martı́n–Mateos, and J.L. Ruiz–Reina

http://www.cs.us.es/{~jalonso,~mjoseh,~fmartin,~jruiz}
Departamento de Ciencias de la Computación e Inteligencia Artificial.
Escuela Técnica Superior de Ingenierı́a Informática, Universidad de Sevilla
Avda. Reina Mercedes, s/n. 41012 Sevilla, Spain
Extended Abstract
We present in this paper a formal proof of the Strong Completeness The-

orem of SLD–resolution, carried out in the PVS verification system. To
the best of our knowledge, the proof presented here is new.
PVS [5] is a specification and verification system integrating an ex-
pressive specification language with an interactive theorem prover and
other supporting tools allowing formal and mechanized reasoning about
the properties of a specification.
The formal proof presented here is part of a complete formalization
of propositional logic programming, carried out in PVS 1 . In that work,
with the purpose of emphasizing a more natural and generic reasoning,
we have choosed finite sets as the main data type used in the definitions
and results of the formalized theory. In this context, we have proved the
equivalence between the different semantics of logic programs. That is,
the least Herbrand model of a logic program P , M P , the set of logical
consequences of the program, LC(P ), and the least fixed point of the
immediate consequence operator, lf p(P ), are the same set:
MP = LC(P ) = lf p(P )
The use of finite sets have allowed the proofs carried out in the au-
tomated reasoning system to have essentially the same declarative flavor
than its corresponding informal mathematical theory presented in most
of textbooks about logic programming. Nevertheless, this correspondence
is no longer possible in the proof of the Strong Completeness Theorem of
?
This research was funded in part by the Spanish Ministry of Education and Science
under grant TIN2004–03884 and Feder funds.
1
In the sequel, when we talk about “proving” some property, we mean that we used
the PVS system to obtain a formal proof of that property.
the SLD–resolution. This theorem states that given a program P , a query
G and a rule of computation f , if P ∪ {G} is not satisfiable, then P ∪ {G}
has a refutation via f (that is, in each step of refutation, it resolves the
atom select by f ). The proof of this theorem is obtained directly applying
the weak completeness theorem (if P ∪ {G} is unsatisfiable, then P ∪ {G}
has a refutation) and the independence of the rule of computation (if
P ∪ {G} has a refutation, then P ∪ {G} has a refutation via f ).
The directed proofs of the strong completeness of the SLD–resolution
made by K. Apt and R. Stärk ([1, 6]) using the notion of implication
tree. On the other hand, the usual proofs of the independence of the
computation rule presented in the literature are based on switching steps
of refutation ([3, 4]). In both cases, they use the implicit notion of finite
sequence, where the order between literals are relevant. Now then, the
switching lemma used by J. Lloyds in [3] to prove the independence of
the rule of computation is not true for a representation of the clauses
based on sets, although the independence of the rule of computation is
still valid, because the underlying idea is that the order of resolution of
the atoms is not relevant.
The proof of the independence of the computation rule is usually made
by induction on the length of the refutation, using the following classic
reduction lemma: if P ∪{G} has a refutation of length n, then for every
atom A of the body of G, there is a clause C in P , such that its head is
A and P ∪ {resolvent(G, C)} has a refutation of length n − 1. The figure
below shows an induction scheme for the proof of the independence of
the computation rule, using the classic reduction lemma to assure the
existence of the clause Cj .
Fig. 1. Induction scheme for the proof.

f (G) = Aj
G :← A1 , ...Aj , ...Am Cj G :← A1 , ...Aj , ...Am
resolvent(G, Cj ) resolvent(G, Cj )
n (via f )
i.h.
n-1 n-1(via f )
This lemma is not true if we consider the bodies of the clauses as
finite sets. Thus, we have generalized this lemma, by defining for each
program P a measure function hP and proving the following generalized
reduction lemma: if P ∪ {G} has a refutation, then for every atom A
of the body of G exists a clause C in P , such that its head is A, P ∪
{resolvent(G, C)} has a refutation and h P (resolvent(G, C)) < hP (G).
Roughly speaking, the function hP measures the “least effort” needed
to prove that an atom or a finite sets of atoms are logical consequences
of P .
In order to define this measure, we take into account that LC(P ) =
lf p(P ) and that lf p(P ) is built using the consequence operator T P , as
the least fixed point of the chain of consequences
TP0 (∅) ⊆ TP1 (∅) ⊆ . . . ⊆ TPk (∅) ⊆ . . .
First, we define the least exponent of an atom A with respect to a
program P . Intuitively, in the case that P |= A, this number represents
the first step in which A appears in the chain of consequences of P :

min{k : k < n ∧ A ∈ TPk (∅)}, if P |= A;
m exponentP (A) =
0, if P |6 = A.
Using it, we define the function hP as follows:
  

  X 

1 + min hP (B) : C ∈ CRP (A) , if P |= A;
hP (A) =  

 B∈body(C)
 n
n , if P |6 = A.
where
CRP (A) = {C : C ∈ P ∧ head(C) = A ∧ body(C) ⊆ T Pk−1 (∅)}
being k = m exponentP (A). If G is a definite goal, we define

X
hP (G) = hP (B)
B∈body(G)
It is worth pointing out that in the definition of h P , when P |= A

and k = m exponentP (A), the least proof of A (in terms of “effort”) has
to use clauses of P whose bodies has been introduced in a chain step
previous to the k-th step (where A was introduced).
This function allows us to prove the generalized reduction lemma
above and, therefore, to prove the independence of the rule of compu-
tation, by strong induction in the measure induced by the function h P .
The formalization made in PVS emphasize the declarative character-
istic of the logic programming, due to both the programs and the clauses
like finite sets. This emphasis places this formalization like basis for the
formal development in PVS of the Answer Set Programming [2].
References
1. K. R. Apt. From Logic Programming to Prolog. Prentice-Hall, Inc., Upper Saddle
River, NJ, 1996.
2. C. Baral. Knowledge Representation, Reasoning and Declarative Problem Solving
with Answer Sets. Cambridge University Press, 2003.
3. J. W. Lloyd. Logic Programming. Springer-Verlag, 2 edition, 1987.
4. A. Nerode and R. A. Shore. Logic for Applications. Springer-Verlag, New York,
NY, 1993.
5. S. Owre, J. Rushby, and S. N. PVS: A propotype of verification system. In
http://pvs.csl.sri.com.
6. R. Stärk. A direct proof for the completeness of SLD-resolution. In E. Börger, H. K.
Büning, and M. M. Richter, editors, Computer Science Logic, selected papers from
CSL ’89, pages 382–383. Springer-Verlag, Lecture Notes in Computer Science 440,
1990.
Fuzzy ontologies generation in GUMSe
Jesus Serrano-Guerrero, Jose A. Olivas, Javier de la Mata
Computer Science Dept.

University of Castilla-La Mancha.
Paseo de la Universidad 4, 13071-Ciudad Real, Spain.
{jesus.serrano3, javierL.Mata}@alu.uclm.es
joseangel.olivas@uclm.es
Abstract
The main target of the ontology construction system here proposed is to solve one of
the main shortages in Internet search engines: “Its nonexistent capacity for semantic
deduction”.
As to endow these so-called semantic capabilities, the idea of automatically gener-
ating ontologies arises, although the majority of the attempts at it until now have failed
due to a basic lack: the definition of ‘processable conceptual relation’. Nowadays, the
main conceptual terms relation measurement is the co-occurrence of such terms as
Widyantoro [1] proposes, which is a main factor but neither determining nor unique.
The main innovation in this study is the redefinition of the idea of ‘processable
conceptual relation’ based on a new influential factor: “the physical distance between
terms”, and based on it, “the differentiation of kinds of conceptual relations”.
Two terms ti, tj , which are in the same sentence, are physically closer than being in
different paragraphs. The explanation of this idea dwells in the fact that the terms ap-
pearing in the same sentence are more bound to be strongly related by predicates with
the form “is a” or “is part of” or even become axioms of the ontology, than two terms in
different paragraphs although partially related to the same matter.
According to this theory, it is possible to find two kinds of different conceptual re-
lations, which does not mean they are independent. Such relations are known as physi-
cal relations and purely semantic relations.
The physical relations are mainly based on both summing up (“is part of”) and gen-
eralization (“is a”) relations since predicates with the form:
The car has wheels
Show in a simple and intuitive way the strong relation between the concept “car”
and the concept “wheel”. This example shows the great importance of the concept:
“physical distance”, for two concepts that are in the same sentence, might have a
strong degree of relationship. On the other hand, purely semantic relations prove that
even in a less intuitive way, some concepts may be related to other ones, as a text
talking about Spanish customs would show. The concept “Spain” would be related to
the concept “bullfighting” -even though Spain is not a bullfighter-, but in a less intui-
tive and therefore ‘weaker’ way. This could also apply to many other customs appear-
ing in different paragraphs, which shows once again that physical distance might be an
influential factor.
Logically, physical relations might also have an important semantic charge at the
same time. In a specific context it is possible to talk about the different types of car
wheels and its importance in rallies and so on, in this context, the relation between
“car” and “wheel” is even stronger, boosted by both physical and semantic relations
between them. Now, it appears the problem of delimitating how high is the degree of
matching between physical and purely semantic relations in two concepts. The relation
between two terms global coefficient is calculated by using the fuzzy intersection to
solve the overlapping zone between physical and purely semantic relation.
Fig. 1. Relational global coefficient.
Relational global coefficient = S f F, where f represents the fuzzy intersection.

The practical proof of this idea has been highlighted in the ontology construction
agent in GUMSe1, a meta-search engine based on an agents architecture developed
within the frame of SMILe 2.
Before a user’s query, GUMSe recovers a number of pages representing the knowl-
edge of the expert arranged according to the semantic capacities given by WordNet [2]
and FIS-CRM [3] conceptual representation model, and from that knowledge it models
an ontology matching its own conceptual relation definition. Such ontology is used in
the phase called “query expansion” of the meta-searcher and in the “user’s feedback”
phase, where the user’s initial query is reformulated so as to increase the initial seman-
tic capacities in the searching phase in GUMSe.
The purpose of GUMSe is to prove that the majority of the current search engines
present a great number of semantic faults that can be partly solved by means of using
ontology created under the idea of identification and separation of two main kinds of
conceptual relations: physical and purely semantic. At the same time, the use of the
“physical distance” factor shows that a great number of documents are not necessary
to generate and refine an ontology, unlike the mere use of the terms co-occurrence, that
needs a large documental basis to achieve acceptable results.
Acknowledg ments
Partially supported by PBC-03-004 PREDACOM project, JCCM, TIC2003-08807-C02-02
DIMOCLUST project, MCYT, and JCCM Cons. de Ciencia y Tecnología - European
Social Fund grants, Spain.
References
1. Widyantoro, D., Yen, J.: Incorporating fuzzy ontology of term relations in a search engine,
in: Proceedings of the BISC Int. Workshop on Fuzzy Logic and the Internet (2001) 155–
160.
2. WORDNET. A lexical database for the English language.
http://www.cogsci.princeton.edu/~wn/ (last visit 4 October 2004)
3. Olivas, J. A., Garcés, P., Romero, F. P.: An application of the FIS-CRM model to the
FISS metasearcher: Using fuzzy synonymy and fuzzy generality for representing concepts
in documents. Int. J. of Approximate Reasoning (Soft Computing in Recognition and
Search) 34 (2003) 201–219.
1 GUMSe: Gum Search.

2 SMILe: Soft Management of Internet e-Laboratory, ORETO-UCLM, Spain.
Extracting Computer Algebra programs from
statements?
Jesús Aransay1, Clemens Ballarin2, and Julio Rubio1

1
Dpto. de Matemáticas y Computación. Univ. de La Rioja. 26004 Logroño (Spain).
{jesus-maria.aransay,julio.rubio}@dmc.unirioja.es
2
Institut für Informatik. Technische Univ. München. D-85748 Garching (Germany).
ballarin@in.tum.de
Davenport stressed in the eighties [4] a difference between Computer Algebra

and Constructive Mathematics: the former is not constrained to a particular kind
of logic (for instance, intuitionism), any proof is suitable if an algorithm can be
designed to compute the sought result from the input. Recently, the interplay
between Computer Algebra and Constructivism has been renewed, as illustrated
by the research initiative MAP (Mathematics, Algorithms and Proofs); see [7].
One of the topics in the MAP project is Sergeraert’s theory for Constructive
Algebraic Topology (CAT). The main realization of the CAT framework is a
Common Lisp program called Kenzo1 ; see [5]. Kenzo is capable of dealing with
highly complex infinite combinatorial spaces (as iterated loop spaces) and can
compute algebraic invariants (as homology groups) which cannot be calculated
by any other means [9]. A question raised at the MAP conferences is to what
extent CAT is a truly constructivist theory. It is obvious that CAT is constructive
in an informal sense (other similar qualifications could be: effective, computable
or algorithmic), since it has finally produced an executable computer algebra
program. However, from a deeper point of view (many of the MAP participants
are specialists in logic or type theory), the question is difficult to grasp. On the
one hand, the previously evoked infinite spaces seem difficult to be modelled into
a weak logic (as the logic in the ACL2 system [6]). On the other hand, proofs in
CAT are explicit and quite “formula handling”, and therefore open to a formal
treatment.
This general question was one of the reasons (the other was of a more prac-
tical nature: to increase the reliability of the Kenzo results) to start a research
line where we use the proof assistant Isabelle [8] to explore properties in CAT.
Due to the intrinsic difficulty of the task we decided to simplify the problem
in two ways: first, restricting our attention to a small (but relevant) fragment
of Kenzo (namely, the BPL or Basic Perturbation Lemma [3]); second, tackling
the problem of the mechanized proof of the theorems, and not of the correctness
of the algorithms (our primary goal). Once the feasibility of this objective has
been established (see [1]), our next goal is to apply the Isabelle code extraction
?
Partially supported by SEUI-MEC, project TIC2002-01626
1
Kenzo is the name of a Sergeraert’s cat. . .
tools [2] to our proofs in order to bridge the gap between theorems and programs
(the extracted code would produce ML programs; another gap would be then to
reach the Kenzo Common Lisp programs).
A simple observation is that many of the theorems in the CAT setting are
of a “constructive” (or “definitional” nature): in the statement a new object is
defined and one property of the new object is asserted. For instance, consider
the case of a monoid (for the sake of simplicity, we use here this quite simple
algebraic structure, instead of the actual data structures from Kenzo, as chain
complexes, simplicial groups, and the like). A theorem claims that if one “twists”
the product of a monoid, obtains again a monoid. Isabelle, as hoped, directly
extract the corresponding ML program.
In Isabelle, this theorem can be expressed as follows:
constdefs rec :: "(’a, ’m) monoid_scheme => ’a monoid"

"rec G == (| carrier = carrier G,
mult = (\<lambda>x y. mult G y x), one = one G |)"
...
lemma (in monoid) shows "monoid (rec G)"
and the (relevant part of the) ML code extracted is
fun rec G =
partial_object_ext (carrier G)
(monoid_ext (fn x => fn y => mult G y x) (one G) Unity);
This seems quite natural, but stresses, in a very concrete way, Davenport’s
claim: if code is extracted directly from the statement it is clear that no trade-
off is acquired with respect to the logic in the proof. However, one could argue
that a “definitional” theorem should be proved in a “property-checking” way, so
obviously in a constructive manner. Being true, this observation does not capture
all the richness of our topic. In general, in CAT, properties are expressed in two
steps: first, a certain construction is proved computable, by constructing it as a
functional between two data types; second, this construction has the convenient
theoretical properties in order to ensure that the results obtained are the right
ones. Let us illustrate this rather speculative claim with another example from
monoids: there is a theorem constructing the product of two monoids. Then,
there is another theorem stating that this object has the right universal property.
It is not difficult (but not mandatory) to give a non-constructive proof of the
uniqueness part of the universal property (by contradiction, with explicit use of
the excluded middle principle). This example proves, in a toy case, that proof
extraction tools can be used in CAT to generate code, but that the underlying
logic could be non-constructive.
Our current research has produced a taxonomy of the lemmas already proved
in Isabelle [1], determining when this “extraction from statements” approach is
enough. More precisely, from the six auxiliary lemmas in Section 2 of [1], three
are directly expressed by constructive statements, and the other three can be
easily transformed into this kind of statement. The only result which seems
essentially different is the BPL itself. Its statement is the following (for the
necessary definitions and notations, we refer to [1]).
Theorem 1. Basic Perturbation Lemma — Let (f, g, h) : D∗ ⇒ C∗ be a

chain complex reduction and δD∗ : D∗ → D∗ a perturbation of the differential
dD∗ satisfying the nilpotency condition with respect to the reduction (f, g, h).
Then a new reduction (f 0 , g 0 , h0 ) : D∗0 ⇒ C∗0 can be obtained where the underlying
graded groups of D∗ and D∗0 (resp. C∗ and C∗0 ) are the same, but the differentials
are perturbed: dD∗0 = dD∗ + δD∗ , dC∗0 = dC∗ + δC∗ , and δC∗ = f φδD∗ g; f 0 = f φ;
P∞
g 0 = (1 − hφδD∗ )g; h0 = hφ, where φ = i=0 (−1)i (δD∗ h)i .
P
The occurrence in the BPL statement of the series φ = ∞ i
i=0 (−1) (δD∗ h)
i
implies further difficulties from a constructive point of view. Some kind of in-
ductive definition would be necessary to deal with the series in a constructive
manner. Then, it should be possible to study if the Isabelle code extraction tool
can be applied in that inductive case.
Another research line is to analyze if the special characteristics of these state-
ments can help to give easier extraction algorithms, or, eventually, to establish
equivalence theorems with ML programs written independently.
References
1. J. Aransay, C. Ballarin and J. Rubio, Four approaches to automated reasoning with
differential algebraic structures, AISC 2004, LNAI vol. 3249, pp. 222-235.
2. S. Berghofer, Program Extraction in Simply-Typed Higher Order Logic, TYPES
2002, LNCS vol. 2646, pp. 21-38.
3. R. Brown, The twisted Eilenberg-Zilber theorem, Celebrazioni Arch. Secolo XX,
Simp. Top. (1967), pp. 34-37.
4. J. M. Davenport, Effective Mathematics: the Computer Algebra viewpoint, Lecture
Notes in Mathematics 873 (1981), pp. 31-43.
5. X. Dousson, F. Sergeraert and Y. Siret, The Kenzo program,
http://www-fourier.ujf-grenoble.fr/~sergerar/Kenzo/
6. M. Kaufmann, P. Manolios and J. Strother Moore. Computer-Aided Reasoning: An
Approach. Kluwer Academic Publishers, 2000.
7. Mathematics, Algorithms, Proofs, http://map.unican.es/
8. T. Nipkow, L. C. Paulson and M. Wenzel, Isabelle/HOL: A proof assistant for higher
order logic, Lecture Notes in Computer Science, 2283, 2002.
9. J. Rubio and F. Sergeraert, Constructive Algebraic Topology, Bulletin des Sciences
Mathématiques 126 (2002) 389-412.
Cross-Language Classification of Function Words
Pablo Gamallo?1 , Gabriel P. Lopes2 , and Joaquim F. Da Silva2

1
Departamento de Língua Espanhola
Universidade de Santiago de Compostela, Espanha
pablogam@usc.es
http://www.sintx.usc.es/ gamallo/
2
CITI, Faculdade de Ciências e Tecnologia
Universidade Nova de Lisboa, Portugal
{jfs,gpl}@di.fct.unl.pt
1 Introduction
Our aim is to describe language-independent properties which will be used to

classify function words across a wide range of languages. These generic properties
are not defined in terms of traditional word distribution, as in most unsuper-
vised methods to learn syntactic categories from raw corpora [2, 5, 6, 1]. Such a
distributional information, which is based on Harry’s Hypothesis, turns to be
inefficient to measure the distance between words across human languages. In
contrast, the features of function words we will describe here are generic dimen-
sions that can be easily used to measure the distance between words belonging to
different languages. In this paper, we will describe features such as, for instance,
the degree of obligatoriness and the degree of dispersion. Regarding obligatori-
ness, each word is assigned a weight representing its tendency to be omitted or
not between two adjacent words [4]. In the scale of obligatoriness, prepositions
and conjunctions are likely to be more obligatory than determiners and auxiliary
verbs. Regarding to the degree of dispersion, a function word can be assigned a
value obtained by computing how sparse are its immediate neighbor words. For
instance, determiners appearing before nouns are assigned low values since the
nouns they modify are likely to be sparse. By contrast, nouns appearing before
prepositions are less sparse because of the syntactic requirements.
So, the main motivation of the article is to discover general principles govern-
ing the distribution of function words across a great variety of natural languages.
The main objective is to use this generic distribution to learn categories of func-
tion words from raw multilingual corpora. The multilingual corpora used in our
experiments is constituted by texts written in three languages, English, French,
and Portuguese. These texts are are not translations of an original one, i.e., the
input of our method is a non-parallel corpus.
.
?
Researcher of the program Isidro Parga Pondal, Xunta de Galicia
2 Cross-Language Features of Function Words
We are interested in discovering cross-language statistical features of the more

frequent function words, in order to learn and induce their basic categories. Our
learning system uses 20 features, but only 1 of them will be briefly outlined in
this abstract. Function words are ranked within a scale of obligatoriness. Given
a function word w, a measure (OBL) that gives us an idea about how much is
w compelled to be part of their contexts, is computed as follows:
1 X f (x, w, y)
OBL(w) = (1)
|(x, y) : ∃(x, w, y)| f (x, w, y) + f (x, y)
(x,y):∃(x,w,y)
where f (x, w, y) denotes the frequency of w in a context constituted by the words,

x and y, immediately before and after w, and where |(x, y) : ∃(x, w, y)| is the
number of unique contexts of w. Moreover, f (x, y) represents the frequency of
word x occurring immediately before y. For instance, preposition “of” cannot be
omitted from an expression like “president of Ireland”, since the bigram “president
Ireland” does not occur in the corpus. Most prepositions have a high value within
the scale of obligatoriness. On the other hand, demonstrative determiners tend
to be optional.
Function words are also provided with information on the degree of sparse-
ness introduced by the words appearing at their right. Following [3], we call
this property “right heterogeneity”. For a given function word w, this measure
(R_HET ER) is obtained by computing a mean of conditional probabilities:
1 X f (w, y)
R_HET ER(w) = (2)
|(y) : ∃(w, y)| f (w)
(y):∃(w,y)
For instance, in many languages, prepositions are often followed by a limited

set of determiners. It means that prepositions are likely to have a quite homo-
geneous context at their right. At the opposite, determiners can appear before
a non restricted set of nouns. Determiners possesses a very heterogeneous right
context.
We also consider left heterogeneity. It measures the degree of word sparseness
we find to the left of function words:
1 X f (x, w)
L_HET ER(w) = (3)
|(x) : ∃(x, w)| f (w)
(x):∃(x,w)
For instance, the past participle “been” has a very homogeneous left context,
since it is followed by a very restricted set of verbal forms: “is”, “has”, “have”, . . .
Features OBL, R_HET ER, and L_HET ER are part of the 20 ones we use
to characterize the vector space of function words. Similarity between each pair
of vectors is computed by the Euclidean distance.
3 Classification of Function Words
We designed a classification algorithm that takes advantage of the knowledge a

speaker has of a particular language. The input can be a very reduced list of
tagged function words. Let’s take an example. In one of our experiments, the
start point for classification was the following list of 5 English function words
and the corresponding POS tags:
PREP with_E
CONJ or_E
ART the_E
DEM this_E
AUX have_E
The tags considered in this experiment were PREP(osition), CONJ(unction),

ART(icle), DEM(onstrative), and AUX(iliar). Given this information and a par-
ticular multilingual corpus constituted by three languages (English _E, French
_F, and Portuguese _P), we obtained the classes plotted in Table 1.
PREP with_E at_E under_E for_E from_E by_E on_E as_E

para_P por_P com_P avec_F pour_F sur_F dans_F
par_F
CONJ or_E and_E e_P ou_P et_F
ART the_E a_E o_P os_P uma_P um_P as_P la_F le_F
les_F l_F
DEM this_E these_E those_E , such_E any _outros_P qual-
quer_P este_P esta_P esse_P estas_P outras_P essa_P
ces_F cette_F ce_F
AUX have_E has_E is_E are_E será_P são_P pode_P
foi_P peut_F est_F a_F
Table 1. Multingual classes of function words.
Then, we split each class in three monolingual word sets. It results in 15 basic
classes of function words (see Table 2).
Each basic class depicted in Table 2 will be used to classify other function
words within a particular language. For instance, the English classes will allow us
to classify rare and ambiguous function words. Such a monolingual classification
will follow a standard method based on context distribution.
ENGLISH
PREP with, at, under, for, from, by, on, as
CONJ or, and
ART the, a
DEM this, these, such, any
AUX have, has, is, are
PORTUGUESE
PREP para, por, com
CONJ e, ou
ART o, os, uma, um, as
DEM outros, qualquer, esta, este, esse, estas, outras, essa
AUX será, são, pode, foi
FRENCH
PREP avec, pour, sur, dans, par
CONJ et
ART la, le, les, l
DEM ces, cette, ce
AUX peut, est, a
Table 2. Monolingual classes of function words.
References
1. A Clark. Combining distributional and morphological information for part of speech
induction. In Proceedings of 10th European Chapter of the Association for Compu-
tational Linguistics, pages 59–66, 2003.
2. S. Finch and N. Chater. Bootstrapping syntactic categories. In 14th Annual Meeting
of the Cogntive Science Society, pages 820–825, 1992.
3. Pascale Fung. Compiling bilingual lexicon entries from a non-parallel english-chinese
corpus. In 14th Annual Meeting of the Cogntive Science Society, pages 173–183,
Boston, Massachusettes, 1995.
4. Pablo Gamallo, Joaquim Ferreira da Silva, and Gabriel P. Lopes. A divide-and-
conquer approach to acquire syntactic categories. In Yasubumi Sakakibara and
Georgios Paliouras, editors, Proceedings of the ICGI’04 Conference, Lecture Notes
in Computer Science. Berlin: Springer-Verlag., 2004.
5. H. Schutze. Part-of-speech induction from scratch. In ACL-93, Ohio State, 1993.
6. T.C. Smith and I.H. Witten. Probability-driven lexical classification: A corpus-based
approach. In PACLING-95, 1995.
Integrating syntactic information by means of
data fusion techniques ?
Francisco J. Ribadas, Jesus Vilares, and Miguel A. Alonso

Dept. de Informatica, Univ. of Vigo Dept. de Computacion, Univ. of A Coru~na
Campus de As Lagoas, s/n, 32004 Campus de Elvi~na s/n, 15071
Orense, Spain La Coru~na, Spain
ribadas@uvigo.es fjvilares,alonsog@udc.es
Extended abstract
In Information Retrieval (IR) systems, the correct representation of a document
through an accurate set of index terms is the basis for obtaining a good perfor-
mance. If we are not able to both extract and weight appropriately the terms
which capture the semantics of the text, this shortcoming will have an eect on
all the subsequent processing.
In this context, one of the major limitations we have to deal with is the
linguistic variation of natural languages, particularly those with more complex
morphologic and syntactic structures than English, as in the case of Spanish.
When managing this type of phenomena, the employment of Natural Language
Processing (NLP) techniques becomes feasible. Previous work has shown that lin-
guistic variation at word level can be eectively managed by taggers-lemmatizers,
improving the average precision attained by stemmers [5]. The next step con-
sists of applying phrase-level analysis techniques to reduce the syntactic variation
present in both documents and queries, identifying the syntactic structure of the
text by means of shallow parsing techniques that provide an eective balance
between descriptive power and computational cost [3]. However, the integration
of the information obtained by syntactic processing (usually in the form of com-
plex index terms) into the indices of an IR system is an open issue. As indicated
in [1], when simple and complex terms are used together, the assumption of term
independence is violated because words forming a complex term also occur in
the documents from which the syntactic dependency has been extracted.
In this article we propose to apply data fusion, a set of techniques for combin-
ing the results retrieved by dierent representations of queries and documents, or
by dierent retrieval techniques, to integrate index terms coming from dierent
text processing techniques. In particular, we consider the fusion of the following
indexing methods:
Single lemmatization. The lemmas of the content words {nouns, adjectives
and verbs, the grammatical categories which concentrate the semantics of a
? The research reported in this article has been partially supported by Min-
isterio de Educacion y Ciencia (MEC) and FEDER (TIN2004-07246-C02-01
and TIN2004-07246-C02-02), by MEC (HF2002-81), and by Xunta de Galicia
(PGIDIT02PXIB30501PR, PGIDIT02SIN01E and PGIDIT03SIN30501PR)
Fig. 1. Weighted mixing of index terms.
text{ are used as index terms. For lemmatizing the documents, we employ
MrTagoo [7], a high-performance part-of-speech tagger based on a second
order Hidden Markov Model, with lemmatization capability, whose input is
provided by a linguistically-motivated preprocessor module [8].
Syntactic dependency pairs. This retrieval method is based on complex
terms derived from the syntactic dependencies existent in the documents
and in the queries. We employ a shallow parser based on a cascade of nite-
state transducers consisting of ve layers [3], whose input is the output of
the tagger-lemmatizer. The result we obtain is a list of syntactic dependency
pairs between the heads of the constituents identied from the text.
Syntactic dependency pairs ltered by relevance feed-back. The index-
ing process is the same described in the previous case, but the querying
process is performed in two stages:
1. The lemmatized query is submitted to the system.
2. The n top documents retrieved by this initial query are employed to
select the most informative dependencies, which are used to expand the
query in order to obtain the nal set of documents. These dependencies
are selected automatically using Rocchio's approach [9] to feedback.
Locality-based retrieval. In this case, we apply a pseudo-syntactic ap-
proach [4] based on the proximity between index terms.We have adapted
the original proposal of Kretser and Moat [6], based on a similarity mea-
sure computed as a function of the distance between terms {in our case
single lemmas{ where each occurrence of a query term has an in uence on
the surrounding terms.
With regard to the combination of the results oered by these retrieval methods,
the following data fusion techniques have been considered:
Weighted mixing of index terms. This fusion method, illustrated in Fig. 1,
merges the sets of index terms extracted from the documents by the dierent
text processing techniques and assigns to each of them its own weights to
be used within the Smart vectorial IR engine [2], which will take them into
account when computing vector distances during the retrieval phase.
In our experiments we have performed exhaustive tests, trying several
weights for each indexing method and also using dierent formulae to com-
pute vector components. The best results were obtained when single terms,
Fig. 2. Reindexing retrieved documents.
extracted by lemmatization, were assigned weights higher than the ones as-
signed to complex index terms.The main problem with this approach is how
to nd the right set of weights, capable of oering a uniform behaviour with
all of the queries being considered, because weights with good results in some
queries do not necessarily maintain their performance with other queries.
Reindexing retrieved documents. In this second method we propose a two-
step approach. As shown in Fig. 2, we perform a rst retrieval process using
single terms extracted by lemmatization. Then, we select the top k ranked
documents and reindex them using complex terms based on dependency
pairs. The nal list of documents is built by reordering the top documents
retrieved in the rst step according to the results of the second phase.
The idea behind this approach is to perform a more precise processing on
the most promising documents and use it to improve the nal ranking. We
have tested this proposal by reindenxing dierent amounts of top ranked
documents and using dierent methods to extract dependency pairs for the
reindexing step. The results obtained were not as good as was expected. The
main reason for this poor behaviour was the high intercorrelation shown by
the index terms extracted during the second step, where the classical formu-
lae to compute vector components does not seen to oer good performances.
List merging based on rank positions and mutual precision. The last
method considered in this paper tries, in contrast with the previous ones,
to perform a fusion of results adapted to each individual query. What
this proposal does is to merge lists of retrieved documents obtained using
dierent indexing methods. As shown in Fig. g-mixing, the nal ranked
list takes into account both the ranking position of each document in the
lists that retrieved it and a value that measures the condence that each
list assigns to those positions.
These condence measures are computed independently for each query and
for each list of retrieved documents, using the precision values obtained for
dierent subsets of the list, taking a combination of the other document lists
as a reference to compute them. As a consequence of this way of computing
Fig. 3. Mutual precision based merging.
these condence measures, the main drawback of this method is that it
improves the ranking position of both relevant and non-relevant documents
that are common to dierent lists. So, this approach can be expected to oer
promising results when dealing with high precision lists of documents.
The corpus employed in our evaluation is the Spanish monolingual corpus
from CLEF (Cross Language Evaluation Forum ), composed by 215,738 news
reports. The set of topics consists of 100 queries, formed by three elds: a brief
title, a one-sentence description, and a more complex narrative specifying the
relevance criteria. All these three elds have been employed to build the running
queries in our experiments, giving a dierent relevance to each of them.
References
1. M. Narita and Y. Ogawa. The use of phrases from query texts in information
retrieval. In ACM SIGIR 2000, pages 318{320, Athens, Greece, 2000.
2. C. Buckley. Implementation of the SMART information retrieval system. Technical
report, Department of Computer Science, Cornell University, 1985.
3. J. Vilares and M. A. Alonso. A grammatical approach to the extraction of index
terms. In Proc. of RANLP 2003, pages 500{504, Borovets, Bulgaria, Sept. 2003.
4. J. Vilares and M. A. Alonso. Dealing with syntactic variation through a locality-
based approach. Lecture Notes in Computer Science (LNCS), 3246:??{??, 2004.
5. J. Vilares, M. A. Alonso, and F. J. Ribadas. COLE experiments at CLEF 2003
Spanish monolingual track. In Working Notes for the CLEF 2003 Workshop, pages
197{206, Trondheim, Norway, Sept. 2003.
6. O. de Kretser and A. Moat. Locality-based information retrieval. In Proc. of 10th
Australasian Database Conference (ADC'99), pages 177-188, Singapore, 1999.
7. J. Grana, M. A. Alonso, and M Vilares. A common solution for tokenization and
part-of-speech tagging: One-pass Viterbi algorithm vs. iterative approaches. in Lec-
ture Notes in Computer Science (LNCS), 2448:310, 2002.
8. F. M. Barcala, J. Vilares, M. A. Alonso, J. Grana, and M. Vilares. Tokenization
and proper noun recognition for information retrieval. In DEXA Workshop 2002,
pages 246250. IEEE Computer Society Press, 2002.
9. J. J. Rocchio. Relevance Feedback in Information Retrieval. In G. Salton, editor,
The SMART Retrieval System{Experiments in Automatic Document Processing.
Prentice Hall, Englewood Clis, NJ, 1971

!"#%$'&() #*)+,.-/0
1 ,)+23)45& ) 67
0%'& ,89&:;,-/0=<>?)&A@"BC%& 34D E
F0GIHKJMLONQPQROSGOTVUXWZYM[INW\R^]_Ga`cbedXWMROfIGIFcdXLOgihjhedOR^kcG lBGnmoJMprqId\stsuJIRvkG HKJ\Lvwud\xOyQP
z*{}|~u~ja/|(~Q(v(c/ZXuX| nt¡£{*¢Z¤vCZo¥Qv¦\~a¡u§X{¨}~|rCBX~(MM\vQ~
r© ² ª;~«( ¬(² ®{¯Q¬Z°?{(±%(² Q³Q/u|*´0|(~(?Mu(0u{}|¹QV~5u|µOZuMu~|c~({3vu¶M}}Cu{3Oº/| ¶M² u%{}/{uv|}({Q~(5Q»^QMQ~I·0MM/{u ¶v·0_¼(/»/|

¸·0_v?3| ² ½(¶v¾|u{~M'Q3M/(¶M·0(/| ¸ ² ~ ² {¿}~·0~j {½Q ¸ u*r»/
~^/u|}{{XX~M}_¶M(~v_V{vM~{¹MM\(¶vO}¾~|}M{¹D |((¶MZ~¾(|(uÀv{~v0D| Q² ~Z={~| ^² u·0{ =¸ »3u}Á\_vrQQ{*v| (² ¶\o{¹¶vQ^}3vt| ;¶vM{}ÂX¶M(
{ ¸² ~(|uu|¶M {| Z² vVQv~/¼|}{¾X|¶M¶MX0u|~| Z² ,3Q~·0³vÃ)Qou|³Ä cQu¿¶M¾\|euQ{{(·n¼ |ÂX~Z~v¼2Å*u=|}{~ (¸ QQuÆÀZ¾|ç {¹Q(
Ç ÈvÉ*Ê^ËOÌ ÍDÎcÏnÊOÐÌ*É
]}L¼p5d\LvÑ½P¹gjwt[IdXwtgiJ\LIP%ÒBN3ÓIdÔ\N3dVhid\stÕMNBÖJ\hihjNZÖ_wugjJML½J\×IwtNØw=ÙIhjNZP³dXLaÚrÒBN3ÒBd\LMw%wuJVsuN wustgiNÔMN
wtÓOJvP¹N2ÚOJÖ[Op¿NLvwuP½suNhiNÔ/dXLvwwuJÛdÜÖ NQs¹w(dXgiL wtJMÝOgiÖ2J\s¼YM[INsuÑ\G)Þ3ÓIgiPrgiP½ßLOJ/Ò LàÒ gwuÓàwtÓIN
LIdXp¿æ N?JX×%]}L^×£J\sup5d/wugjJML'F NwtsugjNQÔ/dXh³á£]FDâVãjäåG
Ptgip¼ÝIhjN5dXÝIÝOstJvd\Ö(ÓKwuJ'wtÓOgePVwudMP¹ßÛÖJ\LIPtgePw(PDJ\×osuNÝOsuNQPtNLvwugjLOÕ2wtÓON5Ú^J^Ö[Op¿NLvwuP0d\LIÚ
wtÖ ÓOJ\Nrp¿YvÝO[O[ONQwtstgiLOÑ,ÕàgjLÜwtÓId5NçPtgiÚ^p¼gePgiw(hedXdXLIsÖÒ3NKdqaÑÛNáÆw×£ÒoJMNQsNgiLèLIPwuw(ÓOdXNçLIÖYvN\[ORaNQqvstÑ'Ñéd¿dXÔ\LINZÚêÖ_wtd\JMhjshcJXwtÓO×%NsuNQÚ^dXhCJ^ÖLv[O[Ip¿prNqnLvNw(s(PPuRoâ3std\NLIwt[OÚ'suwtLOÓOgiLONQLÕ
wtJKwuÓONë[IPtNsrwtÓON2ÖhjJvP¹NZPwJMLONQP2ã ìåGíêÓONL wuÓON2L[Op½qaNQsJX×Ú^J^Ö[Op¿NLvwuPrgiPrhid\stÕMN\RwuÓOgiP
dXÝOÝOsuJMdMÖ(Ó2suNQYv[OgistNZP3d½hiJXw J\×%ÖJ\p¿ÝO[^w(d/wtgiJ\L,wtgip¿N\G
dXLIÚK]}LÖ(ÓOJMJsuJvÚOP¹NNrs0dëwuJKstNQstÝONZstÚ^NZ[IP¹NQÖ LvNëwud/PtJ\wugjp¿ÔMN?N5JXJ\×=×BNQwudMÓOÖ(NQÓKPtN5stNZJ/P¹Ô\[ONQhjstwtÓIgiLONQÕëd\ÚIÕMP?stJMJ\[OLOÝîNëGIÖQÞ3dXÓILàNLîÖ hiRa[IwtP¹ÓOwtNN¼sYvwt[OÓONN2suÑ2ÚOJÓIÖdM[OP p¿wuJ2NLvqnwuNP
p¿dXwuÖ(ÓONZÚ'J\LOhiÑ,d\ÕMdXgiLIP¹w wtÓON0suNÝIstNZP¹NQLMw(d/wugjÔMNQPQROdXLIÚ'LOJ\wcdXÕMd\gjLaPw wuÓONÒ ÓOJMhjNÖJ\hihjNZÖ_wtgiJ\LCG
]¾×wtÓON Ö hi[IP¹wtNQstgiLOÕïd\hjÕMJ\sugwuÓOp.geP,LOJ\LðP¹[OÝnNsuÔgiPtNQÚîR*J\LINçÖd\LñdÔ\JMgiÚèÓ[OÕMNÛNònJMs¹w(Pë×£JMs
hid\qaNQhjhigjLIÕ0wtÓONcwts(dXgiLOgiLOÕrPtN wQGM]}L2d\ÚIÚ^gwugjJMLîRXwuÓONQPtNDd\hjÕMJ\sugwuÓOp5P³PtÓIdXhihIqnNcÚ^NZP¹giÕ\LONZÚ¼wtJrwudXßMN
d\Ú^Ô/dXLvw(dXÕ\NVJX×%ÝId\sud\hjhiNhdXs(Ö(ÓOgjwtNZÖ_wt[IstNZPoÒ ÓIgiÖ(ÓóÖd\L'gip¿ÝOsuJ/Ô\NDwtÓONQgjs Nô5ÖgjNQLIÖ ÑMG
õ öçËO÷øÐÌ*ÎDù ú Ì³ËOû
Þ3ÓON=wÑÝOgeÖdXhM[IPtN³JX×IüNQh×Æý¹þDstÕvdXLOgiWgiLOÕScN wÒBJ\sußP;giL]F gePîwuJcÖ hi[IP¹wtNQs)Ptgjp¿gihid\sCÚ^J^Ö [Op¿NQLMw(PG
wtÞ3ÓOÓONQNëp¿P¹PtNZNdXhiÔ\s(NZÖ(ÓçPRvgesuP0NQÚ^wu[IÓOÖNgjLàLIÕ0ÝnNwtÓOst×£NQJ\L5sup¿wtÓINQNVÚçÖJ/J\Ô\p¿NQs?ÝO[^wtÓOw(d/NëwtgiLIJ\NLa[OdXsuhaJ\ÖLIJ\P?p¿giÝOLIhiP¹N wtØ^NZd\gjwÚÑ\GJXÞ3× ÓOJ/NÔ\suNQNsVgeP*wtÓONQN2ÔvgeÚOÚ^JNQÖLI[OÖ p¿NcwtNÓILvdXwuwP
wtÓOgePîp¿giÕ\ÓvwîhjNZd\ÚVwuJdXL?gjp¿ÝOsuJ/Ô\NQp¿NLvwgjL?wuÓON*Ú^J^Ö [Ip¼NQLvwCstNwtsugjNQÔ/dXh^ã ÿ/åvdXwîhjNZd\P¹wCÖ JMp¼ÝadXsuNQÚ
Ò gwuÓóPtgip¼ÝIhjN0]Fðp¿N wuÓOJ^ÚOP P¹[aÖ(ÓódMP ü^dXhjwtJML P3T=NZÖ_wuJ\sDüÝad\Ö N0HKJ^Ú^NQh³á£T?ü^HÛâ?ãìå¾G
±%² c{(}_{ ² {u^Q{}|_ë~2| ² ~ vQû{ ² QB»^((2}¶vv^{}|_5~,v{}| ¶MMt { é{¹Q/|
±
Q
Þ3ÓONsuNrNØgeP¹wuP?P¹JMp¿N½ÒBJ\suß^PcÒ gjwtÓçstNQÕMdXs(ÚówuJ2wuÓON5Ú^NQPtgjÕMLÛJ\×oÖ hi[IP¹wtNsugiLOÕ'd\hjÕMJ\sugwuÓOp5P
qId\PtNQÚ½J\L¿ÝId\sud\hjhiNh^dXs(Ö(ÓOgjwtNZÖ_wt[IstNZP%Ò ÓOgiÖ(Ó¿Öd\L¿gjp¿ÝOsuJ/Ô\NowtÓON NôëÖ giNLIÖÑrJX×dXL¼]FèP¹Ñ^P¹wtNQp'G
]}d LïÙOã Ø^/å³NQÚèdëÝOÝIsud\J\suqId\hjhjhiNQNp'h=GoüIþV]}LñH d\ÚIgePVÚ^qOgwu[OgjJMgihLêwZRngp5w,d\giP'ßvgiÚ^LONQÕ,PtgiN Õ\Ø^LOÝaNZNQÚéstgip¿d NüILvþVwuPDH J/Ô\Ò NQsDÓOged\Ö(LÜÓ] gi3P,H dXÝOüÝIohjgiNQì,ÚèdXLIÒ ÚÛÓONQÒ LgwugjÓ w
ÒoJMstß^PVÒ gwuÓ!P¹ÝadXs(P¹N½wts(dXgiLOgiLOÕóN ØOdXp¿ÝOhiNQPQGCüJ\p¿N5J\wtÓONQs0ÒoJMstß^P ã åogjLvwtsuJ^Ú^[IÖN¼ÝadXs(dXhihjNQh
Ö J\p¿ÝO[OwtgiLOÕ¼wtNQÖ(ÓILOgiYv[ONZPowtJ¿Ú^J^Ö [Op¿NLvw giLIÚ^N Ø^giLOÕIR^p5d\ßvgiLOÕ¼ÓOgiÕ\ÓONQsowuÓON?ÝaNQs¹×£JMstp5d\LIÖ NDJX×
Ú^gò;NsuNLvwcÖhidMPtPtgiÖQdXh;suN wtsugiNÔ/dXhdXhiÕ\JMstgjwtÓIp¿PQG
ö %Ë !"÷#$èËa&Ï %cÐÊ^÷CÏnÊOÎËO(÷ '(Ì³)Ë =+ É *ÉDùvÎ ,0÷îË^øÐ¹ùv÷CÍ .È -
/,0,cËO1 Ì =2Ï %
Þ3ÓONrNØgeP¹wtNLaÖ NrJ\×³ÓIdXs(Ú^Ò3dXsuN ý}P¹J\×ÆwÒBd\stNDÝIdXs(dXhihiNh)PtJ\hi[^wugjJMLIPJ\×³hjJ/Ò Ö JvPw?dXLIÚÜÓOgjÕMÓKÝaNQs¹ý
×£þ?J\üIsup¿m d\æ LIFcÖ N'â2dXá dMhihjPcJ/Ò×£JMPës?[IN ØOPëdXwup¿J ÝOP¹hiwtN½[I3d ÚOÑèBm dïLIP¹JMN hjw[^ÒBwuJ\gjJMsuLêß,[IÒ PtÓOgiLOgeÖ(Õ'Ód2gjL×£suÔMNJ\N¼hiÔ\J\NQÝnPëN[OL^LIý}P¹PtJM[O[OÝas(NQÖ stNÔhigegjP¹ßMNZ5NÚè4^hjNZÞdX6ü suLOgiLOJMÕ s
dXLIÚ,ÝIdXs(dXhihiNhÖ JMp¼ÝI[^wtgiLOÕ5Ò gjwtÓód½ÒBNhihîÚ^NÙILONQÚ'ÕMJMdXh 7^Þ3ÓON0]FðP¹Ñ^P¹wtNp pr[IP¹wqaN?ÓOgiÕ\ÓIhjÑ
×£PtJ\ÖQsudXÕ\hedXN qOwtwthigiN\LORîÕ¼NôëwtJ¿Ö gigiNp¼LvÝIwQstRJ/Ö(ÔMÓONcNQd\wuÓOÝaN0NQs0J/Ô\wtÓINQsud\d\L hjh;wt]s(F d\Ú^ÝOgjwtstgiJ^J\ÖLaNQdXPuh³PG JMLONQP½dXLIÚhiJ/Ò3ýp5dXgiLvwtNQLIdXLIÖN5Ò gwuÓOJ\[Ow
á d\P¼]}×£L2JMs¿× dMN Ö_ØOwQdXRap¿P¹JMÝOp¼hiNóN0üaÖhjþV[IHA P¹wtNQstã 8/giLOå£â¼Õ5dXdXsuhiÕ\N2JMÙIstgjs(wtP¹ÓIw5p¿P¹PBwt[aqIÚ^dMgjP¹NZNZÚÚ2G JMæ LK×ÆwuP¹NNQs¿h×Æwuý¾ÓIJ\sud/ÕMwZd\R=LOwtgjJ WQgjLIPwuÕr[ILOÚ^ÑàNQ[OwusuÓOd\hNKLON N ò;wNQÒBÖ J\w5suß^JXP×
ÝIdXs(dXhihjNQhBd\suÖ(ÓIgwuNQÖ_wu[OsuNQPJ/ÔMNs½wtÓON']}L^×£JMstp5d/wugjJMLïF NwtsugjNQÔd\h P¹Ñ^P¹wtNp5PQRwtÓONZP¹NóÖ hi[IP¹wtNsugiLOÕ
dXhiÕ\J\sugjwtÓOp5Prd\stN,Ö J\p¿ÝId\stNZÚçwtJÜwuÓONgis¼ÝIdXs(dXhihjNQh*ÔMNs(P¹giJ\Làd\P½ÒBNhihod\P½Ò gjwtÓïdçP¹gip¿ÝOhjN,]F
p¼NwtÓOJ^ÚñLIdXp¿NQÚTVü^H ã ì/åG Þ3ÓONQstN×£J\suN\R³wuÓON Ö hi[IPwuNsugjLIÕ dXhiÕ\JMstgjwtÓIp¿P,d\P,ÒoNQhjh?d\P2wtÓIN
wts(d\Ú^gjwtgiJ\LId\h*dXÝIÝOstJvd\Ö(ÓONZP?dXsuN5ÝIdXs(dXhihjNQhjgiWNZÚqvÑÚ^gePwustgiqO[^wugjLOÕÜwtÓON,N ØOdXp¿ÝOhiNQP0giL!PtNÔ\NQsud\h
ÒBdÑ^ÞCPQJ,G wtNZPw0wuÓON5ÝaNQs¹×£JMstp5d\LIÖ N¼JX×BJ\[Os?p¿N wuÓOJ^ÚOP0gjw?geP?[IPtNQÚçwtÓI:N 9<;>=@?BADCEGFDEIHKJàã LåGîÞ3ÓIN
ÝaNQs¹×£JMstp5dXLaÖ NrJ\×*wuÓON½ÝadXs(dXhihjNQh%PtÑ^PwuNp5PVÒ gjhih%qaN¿ÒBNgiÕ\ÓvwtNZÚÜgiLÜwuNsup¿PVJ\×*wtÓIMN AN/COEIAJ@CP
dXLIÚ'wuÓOQN FCDRTSURWVOEXSYEXZPJ\×=wtÓON½P¹Ñ^P¹wtNpóGnÞ3ÓONrqnNÓIdÔgiJ\sJ\×%wtÓON¼P¹Ñ^P¹wtNp5P ÝIstNZP¹NQLMwuNQÚógiLKwtÓIN
ÒoJMstß5gePÖ JMp¿ÝIdXsuNQÚRv×£JMs N ØOdXp¿ÝOhiN\RwuJ¿wtÓOJvP¹N?J\qOwudXgiLONZÚ2qÑëT?ü^H J\scüp5dXstwQG
-é[ ÷ 'u÷CËO÷CÉcÏ;÷)ù
\ À\Q|ZI ]DÀO ¼O ]B~Æ ,À ^M`À _Xzo ~/|}{XM¶Mu|~Z3|³·0XMt{ ~\e{·|~Z {u|}{~(QQÆKÀ ¼a ]o{¹ ¸
b ~[ c \ d ef
/À\Q|ZI1]DÀhg'Zv\%zÀ#iQMM%oÀY_ézuu|Q{r}(5·0XMu*iQ{¼Q¶M|Q·|~5~\uÁ\~MMÀ
eXq À papIuQQ|}Z·0|¶Mu·V({@q¶vr?Mj.~t$`À __ cI±| ~uZÁXM|3 =fu{Qt |}\ {| o.~² u Z z*C! ¶Mk}~vjK¼l$}uc \ jQm{ f1vn \`~ s(eO_oK'n M /(¶M·0(/|3·QM(À B(¶\{¹QCÃî{/u(}}~v
\q À }pa( ¸j{u{v(vZa`~ s(Å~vÀ rvD·ÀaQz*M·}~|~Æ2² ]DÀv^vÀn~_Å* ¶v|~ Q² v·0*(|~?u{_} b {À =}V`À _0Mz|¹V}_·0Q~Qv»M~M0?M{{¹Q»v(u·0(wÀ QQv3{~| |¹² < · ¼i~Qa{ À
r M ¸ ÆÀxv3~}( OÀy3c \ f \ m \ o \
¸
/z À {¿u|}{a ~{(QQ{I¾v\Z¾n|(zc· À@ ¶v¼}~ M%} Q¿MaÃC2 ¹^MÀCÀ zc2 À_MaÃ Åo~qQ»^:u{}¿|}ZÃCI¹aÀ Oc \ À = YÀ _? f#Ã e pI\ | maîoTÅ3e O6 _Izñv{¹(î~MiQ{· |~Q
nXÀ
ro ² Qv(I ± ÀY_¼X(io{v~`s(~vQ¿v(À@~)Q¶M·0e¼Q*XM{~vQu{V\u{~(~'MiQ{·|~Q
m/
À ~CÃX({/~/Q(({v(² (_(Q\}À%~(vXT Z=M*{À ,~QvQÀ|u²b1{ q~C{îu·{~Q² a | ²5cIvV ± À r?u ÁX\ `À |³_^f z*Å*vu|}^{(~(vZM~Áa_ îZQMQiu¶{( M|~(Q}Vc |± (Å ² )M~jº/¶Md(ftCoQ}c ± ·0cI_ Q} ¶\{(f (À2 _
An Iterative Method for Mining Frequent
Temporal Patterns
Francisco Guil1 , Alfonso Bosch1 , Antonio Bailón2 , and Roque Marı́n3

1
Departamento de Lenguajes y Computación
Universidad de Almerı́a
04120 Almerı́a
{fguil, abosch}@ual.es
2
Dept. Ciencias de la Computación e Inteligencia Artificial
Universidad de Granada
18071 Granada
bailon@decsai.ugr.es
3
Dept. Ingenierı́a de la Información y las Comunicaciones
Universidad de Murcia
30071 Espinardo (Murcia)
roque@dif.um.es
Abstract. The incorporation of temporal semantics into the traditional

data mining techniques has caused the creation of a new area called
Temporal Data Mining. This incorporation is especially necessary if we
want to extract useful knowledge from dynamic domains, which are time-
varying in nature. However, this process is computationally complex,
and therefore it poses more challenges on efficient processing that non-
temporal techniques. Based in the inter-transactional framework, in [2]
we proposed an algorithm named T SET for mining temporal patterns
(sequences) from datasets which uses a unique tree-based structure for
storing all frequent patterns discovered in the mining process. However,
in each data mining process, the algorithm must generate the whole
structure from scratch. In this work, we propose an extension which
consists in the reusing of structures generated in previous data mining
process in order to reduce the time execution of the algorithm.
1 Introduction
Data mining is an essential step in the process of knowledge discovery in da-
tabases that consists of applying data analysis and discovery algorithms that
produce a particular enumeration of structures over the data. There are two
types of structures: models and patterns. So, we can talk about local and global
methods in data mining. In the case of local methods, the simplest case of pattern
discovery is finding association rules [1]. The initial motivation for association
rules was to aid in the analysis of large transactional databases. The discovery
of association rules can potentially aid decision making within organizations.
Another approach is integrating the data mining process into the development
of Knowledge Based Systems.
Since the problem of mining association rules was introduced by Agrawal
in [1], a large amount of work has been done in several directions, including
improvement of the Apriori algorithm, mining generalized, multi-level, or quan-
titative association rules, mining weighted association rules, fuzzy association
rules mining, constraint-based rule mining, efficient long patterns mining, main-
tenance of the discovered association rules, etc. We want to point out the work
in which a new type of association rules was introduced, the inter-transaction
association rules [3]. Temporal data mining can be viewed as an extension of
this work.
Temporal data mining can be defined as the activity of looking for inte-
resting correlations or patterns in large sets of temporal data accumulated for
other purposes. It has the capability of mining activity, inferring associations of
contextual and temporal proximity, some of which may also indicate a cause-
effect association. This important kind of knowledge can be overlooked when the
temporal component is ignored or treated as a simple numeric attribute [4].
We can find in the literature a large quantity of temporal data mining tech-
niques. We want to highlight some of the most representative ones. So, we can
talk about sequential pattern mining, episodes in event sequences, temporal as-
sociation rules mining, discovering calendar-based temporal association rules,
patterns with multiple granularities mining, and cyclic association rules mi-
ning. However, there is an important form of temporal associations which are
useful but could not be discovered with this techniques. These are the inter-
transaction associations presented in [3]. The introduction of this type of asso-
ciations was motivated by the observation that many real-world associations
happen under certain context, such as time, place, etc. In the case of temporal
context, inter-transactional associations represents associations amongst items
along the dimension of time. Due to the number of potential association be-
comes extremely large, the mining of inter-transaction association poses more
challenges on efficient processing than classical approaches. In order to make the
mining of inter-transaction associations practical and computationally tractable,
several methods have been proposed in [5].
Working in the same direction, in [2] we presented an algorithm named T SET

based on the inter-transactional framework for mining frequent sequences (also
called frequent temporal patterns or frequent temporal associations) from several
kind of datasets. The improvement of the proposed solution was the use of a
unique structure to store all frequent sequences. The data structure used is the
well-known set-enumeration tree, commonly used in the data mining area, in
which the temporal semantic is incorporated. Although the proposed solution
requires less computational resources than other approaches, mining this sort of
associations is still a computationally intensive problem.
In order to reduce the time execution of the algorithm, in this paper we pro-
pose an extension of T SET which is based in the reusing of structures generated
by previous data mining process.
2 An Iterative Mining Process
Knowledge Discovery and therefore data mining is a human-centered process.

After setting the desired values for the user-defined parameters of the selected
algorithm, the user analyze the result of the data mining step in order to extract
useful knowledge. In particular, T SET has two user-defined parameters, mini-
mum support (minsup) and the length of the temporal windows, (maxspan).
Generally, after setting the value for maxspan, the user carries out an incremen-
tal knowledge discovery process modifying the value of minsup in a gradual way.
The incremental process consist in the reduction of the minimum support and
the study of the new discovered patterns. However, as we can see in [2], with the
reduction of this value the number of discovered temporal associations and the
execution time of the algorithm increases exponentially. In order to avoid this
exponential trend, we propose an iterative process that consist in the reusing of
the structure generated by previous data mining process (with higher minsup
values).
We will use an example to illustrate the basic idea. Suppose that Figure 1
shows the data structure generated by T SET from a dataset D, with maxspan =
1 and minsup = 2.
Fig. 1. The initial extended set-enumeration tree structure
If the user sets the minsup value to 1, instead of generating the whole struc-
ture starting from the scratch, the algorithm iterates over the structure looking
for nodes with support less than or equal to the chosen value. In Figure 2 we
can see the generated structure.
To check the usefulness of this approach, an empirical evaluation is being
carried out over several kind of datasets. In particular, we are studying the algo-
rithm with real databases which contain phytosanitary treatment information.
Fig. 2. The second extended set-enumeration tree structure
References
1. R. Agrawal, T. Imielinski, and A. N. Swami. Mining association rules between sets
of items in large databases. In P. Buneman and S. Jajodia, editors, Proc. of the
ACM SIGMOD Int. Conf. on Management of Data, Washington, D.C., May 26-28,
1993, pages 207–216. ACM Press, 1993.
2. F. Guil, A. Bosch, and R. Marı́n. TSET: An algorithm for mining frequent temporal
patterns. In Proc. of the First Int. Workshop on Knowledge Discovery in Data
Streams, in conjunction with ECML/PKDD 2004, pages 65–74, 2004.
3. H. Lu, L. Feng, and J. Han. Beyond intra-transaction association analysis: Mining
multi-dimensional inter-transaction association rules. ACM Transactions on Infor-
mation Systems (TOIS), 18(4):423–454, 2000.
4. J. F. Roddick and M. Spiliopoulou. A survey of temporal knowledge discovery
paradigms and methods. IEEE Transactions on Knowledge and Data Engineering,
14(4):750–767, 2002.
5. A. K. H. Tung, H. Lu, J. Han, and L. Feng. Efficient mining of intertransaction asso-
ciation rules. IEEE Transactions on Knowledge and Data Engineering, 15(1):43–56,
2003.
6. Zhi Hua Zhou. Three perspectives of data mining (book review). Artificial Intelli-
gence, 143:139–146, 2003.
Data Mining with Scatter Search?
I.J. Garcı́a del Amo, F. Garcı́a López, M. Garcı́a Torres, B. Melián Batista,
J.A. Moreno Pérez, J.M. Moreno Vega and Raquel Rivero Martn
Dep. de Sistemas Informticos y Computacin, Universidad de La Laguna,

38271 La Laguna, Spain
1 Introduction
Processing Intelligent Information requires efficient tools to extract the useful in-
formation stored in databases. Data Mining and Knowledge Discovery are power-
ful technologies for the extraction of information from large databases. Heuristic
approaches are becoming very useful for designing efficient tools in Data Mining
[1]. Most of data mining tasks are solved by the application of Machine Learning
techniques. Three central paradigms for the application of Machine Learning [9]
in Data Mining are Clustering, Instance-Based classification and Feature Selec-
tion. We describe the main components of the Scatter Search Metaheuristic and
their specific design to solve standard problems in these tasks. This method has
been tested for these or similar problems ([3], [2]).
Scatter Search [8] is an evolutionary population-based metaheuristic that
constructs solutions by combining others in an evolving set of solutions, named
reference set (Ref Set). The two main differences between Scatter Search and
other classical population-based evolutionary procedures in Data Mining [7] like
Genetic Algorithms [6] are the size of the evolving set of solutions and the way
the method combines the existing solutions to provide new ones. The evolving
set Ref Set has relatively small or moderate size (typical sizes are 10 or 15, see
[8]). Scatter Search combines good solutions to construct others exploiting the
knowledge of the problem at hand in an intelligent way. Genetic Algorithms are
also evolutionary algorithms in which a population of solutions evolves by using
the mutation and crossover operators. The mutation and crossover operators
used in Genetic Algorithms have a significant reliance on randomization to create
new solutions.
2 Scatter Search Metehauristic
In a Scatter Search algorithm [8], a moderate sized set of solutions, the reference
set Ref Set, evolves due to mechanisms of intelligent combination of solutions.
This RefSet is generated and iteratively updated attempting to intensify and
diversify the search. After intelligently combining the solutions in the reference
set, a local search procedure is applied to improve the resulting solution, and
?
This research has been partially supported by the Ministerio de Ciencia y Teconologı́a
through the project TIC2002-04242-C03-01; 70% of which are FEDER funds.
the RefSet is updated to incorporate both good and disperse solutions. These
steps are repeated until a stopping condition is met. The method provides not
only a single heuristic solutions, like other metaheuristics, but a reduced set of
disperse high quality solutions.
The principles of the Scatter Search metaheuristic were first introduced in the
1970s as an extension of formulations for combining decision rules and problem
constraints. This initial proposal generates solutions taking account of charac-
teristics in several parts of the solution space [4]. Scatter Search includes five
main components processes: 1) Diversification Generation Method, that gener-
ates a set of diverse solutions, 2) Improvement Method, that improves a solution
to reach a better solution, 3) Reference Set Update Method, which builds and
updates the reference set consisting of Ref SetSize good solutions, 4) Subset
Generation Method, to produce subsets of solutions of the reference set, and 5)
Solution Combination Method, that combines the solutions in the produced sub-
sets. A comprehensive description of the fundamentals of Scatter Search can be
found in [5].
procedure Scatter Search

begin
Diversification Generation Method;
Improvement Method;
repeat
Reference Set Update Method;
Subset Generation Method;
Solution Combination Method;
Improvement Method;
until (StoppingCriterion);
end.
Fig. 1. A Scatter Search Metaheuristic Pseudocode
A basic Scatter Search method is shown in Figure 1. The algorithm starts

generating a population of solutions by the Diversification Generation Method.
This procedure creates a large set of diverse solutions that are improved by the
Improvement method. A representative set of good solutions are chosen to be in
the reference set (Ref Set). At each iteration of the Scatter Search several subsets
of solutions from the Ref Set are then systematically selected by the Subset
Generation Method. The Solution Combination Method combines the solutions in
each subset taking account their good features. Then, the Improvement Method is
applied to the result of the combination to get an improved solution. Finally, the
Reference Set Update Method uses the obtained solution to update the reference
set following intensification and diversification criteria.
3 Application of Scatter Search in Data Mining
Three main paradigms in Machine Learning applications in Data Mining are:
clustering, instance-based learning and feature selection. We describe the specific
design of the main components of the Scatter Search to solve standard problems
in these three tasks.
Clustering is the main paradigm of unsupervised learning. The objective of
clustering is to find groups of instance constituted by similar instances. Given a
set of instances described by a series of features, the problem is to find a partition
of the whole set in subsets or classes in such a way that instances in the same
class are very similar and instances in different classes are very dissimilar. The
distance based approach considers a distance between instance descriptions to
evaluate the similarity and dissimilarity. A very usual way to define the partition
consists of finding some representative instances for the classes (in the simplest
case only one instance is chosen for each class). Then each instance is assigned
to the class of the nearest of these representative instances. Scatter Search has
been successfully applied to the p-median location problem that is very similar
to the Clustering Problem [3]. The p-median problem consists in choosing the p
points that minimize the sum of distances to the remainder instances.
In instance based supervised learning we consider, in addition to the features
that describe the instances, an additional variable that represents the class of
the instance that is to be predicted from its description. From a training set
of instances with known classes we want to get a classification rule to obtain
the unknown classes of a set of test instances. The distance based classification
approach also selects a set of representative instances from the training set and
classifies the test instances taking into account the class of the nearest selected
instance. This problem is also similar to the p-median problem since it also
consists of selecting a number of instances with a different optimization function.
They belong to the wide set of the named p-selection problems for which most of
the heuristic procedures are based on interchanges in the solutions. The scatter
search approach for a p-selection problem based on interchange moves can be
easily adapted to other problem in this class.
However, the use of the whole set of features is not useful for being con-
sidered in this or other classification paradigms. The feature selection problems
tries to get the best subset of features to perform the classification task. The
appropriated selection of features has not only the advantage of taking the rele-
vant information in the description of the instances, but also avoiding redundant
information and making the classification algorithms and rules more efficient to
obtain and to use. Scatter Search has been tested for this problem in [2].
A distance function between solutions plays a central role in the Scatter
Search to modulate the diversification and intensification. Given a distance be-
tween the items that constitute the solutions (instances for clustering and clas-
sification and variables for feature selection), the distance between two solutions
is the sum of the distances between the items in one solution and the other so-
lution. Similarly, the most diverse solution with respect to a set of solutions is
defined in similar way.
The Reference Set Update Method generates and updates the reference set by
quality and diversity criteria. Solutions for the reference set are first chosen by
quality; e.g. the Ref SetSize/2 best solutions. Then the most diverse solution is
chosen Ref SetSize/2 times.
The usual Subset Generation Method in the applications of Scatter Search
consists of considering all the subsets of a fixed size (usually two) of solutions
in the current reference set of solutions. The solutions in the subsets are then
combined to construct other solutions avoiding repetitions if the subset have been
previously used in a combination. The Solution Combination Method combines
good characteristics of the selected solutions to get new current solutions.
The combination methods for these problems are random/greedy strategies.
They start with a partial solution consisting of the items common to the solution
to be combined. Then, at each iteration, one of the remaining items in some of the
combined solutions is added. The criteria applied to select the items are between
the pure random and greedy criteria and consists of selecting at random one of
the most improving item.
The Improvement Method applied to the solutions of the population and those
generated by the combination method are usual local searches. They are mostly
based on the basic interchange method that consisting in replacing an item in
the solution by an item out of the solution. The solutions obtained by improving
the combined solutions are recorded in a pool of solutions, ImpSolSet, which
is then used to update Ref Set by the Reference Set Update Method using the
above method applying intensity and diversity criteria.
References
1. H.A. Abbass, C.S. Newton , R. Sarker. Data Mining: A heuristic Approach. Idea
Group (2002).
2. F. Garcı́a López, M. Garcı́a Torres, B. Melián Batista, J.A. Moreno Pérez and
J.M. Moreno Vega. Solving Feature Subset Selection Problem by a Parallel Scatter
Search, European Journal of Operational Research, 2005, to appear.
3. F. Garcı́a López, B. Melián Batista, J.A. Moreno Pérez and J.M. Moreno Vega.
Parallelization of the Scatter Search for the p-median problem, Parallel Computing,
29 (2003) 575-589.
4. Glover, F., Heuristics for Integer Programming using Surrogate Constraints, Deci-
sion Sciences 8, (1977) 156–166
5. Glover, F., Laguna, M., Martı́, R. Fundamentals of Scatter Search and Path Relink-
ing Control and Cybernetics, 39, (2000) 653-684
6. Goldberg, D.E., Genetic Algorithms in Search, Optimization and Machine Learning,
Addison Wesley, (1989)
7. Ghosh, A. and Jain, L.C. (Eds.) Evolutionary Computation in Data Mining Studies
in Fuzziness and Soft Computing, 163. Springer (2005)
8. Laguna, M. and R. Martı́, Scatter Search: Metodology and Implementations in C,
Kluwer Academic Press, (2003).
9. Mitchell, T. Machine Learning, Series in Computer Science, McGraw-Hill, (1997).
This article was processed using the LATEX macro package with LLNCS style
Web Usage Mining Project for Improving Web-based
Learning Sites
M. E. Zorrilla1, E. Menasalvas2, D. Marín1, E. Mora1, J. Segovia2

1
Department of Applied Mathematics and Computer Sciences, University of Cantabria.
Avda. de los Castros s/n 39005 Santander. Spain
{zorrilm, morae}@unican.es, dmoujo@yahoo.es
2
Dpto de Lenguajes y Sistemas Informáticos. Facultad de Informática
Universidad Politécnica de Madrid. Boadilla del Monte. Spain
{fsegovia, emenasalvas}@fi.upm.es
Extended Abstract
Relationships with the users are paramount when trying to competitively develop ac-
tivities in any web environment. Information collected and stored by Web servers is
the main source of data for analyzing user navigation patterns. A successful e-CRM
depends mainly on the capabilities of the organization to identify and classify users in
order to model and predict her/his behaviour on the site. Web site personalization is
one of the keys to do it [2]. Different web mining approaches and techniques have
been proposed in order to improve the sites by creating adaptive Web sites, based
mainly on the analysis of user access logs [5].
Higher education (HE) is an example of an activity that is having success when be-
ing developed through the web. Students and professor are provided with easy to use
tools and are not subject to any temporal constraint. However, it is very difficult and
time consuming for educators to thoroughly track and assess all the activities per-
formed by all learners on all these tools. Moreover, it is hard to evaluate the structure
of the course content and its effectiveness on the learning process [8].
E-learning environments lack a closer student-educator relationship. The lack of
this relation is manifested in facts such as: teacher does not really control the evolu-
tion of his students, and students cannot express their problems and deficiencies in a
natural way.
The roots of this problem lays in web-based learning environments, such as Vir-
tual-U [3] and WebCT [6], including course content delivery tools, synchronous and
asynchronous conferencing systems, polling and quiz modules, virtual workspaces for
sharing resources, white boards and other tracking tools but not including specific
tools to track and analyze student behaviour and their evolution.
In fact, these environments provide the educator with access summary information
such as more visited pages, favourite communication method, and other statistics.
Nevertheless this information is not enough to analyze the behaviour of each student
when the number of students increases.
Although data mining has been successfully applied in e-commerce sites to im-
prove the lost 1-to-1 relationship, it is not yet engaged into e-learning systems. Never-
theless, web-based learning systems can profit from the integration of data mining and
the e-learning technology [4].
The present paper presents advances of an e-learning project in which data mining
techniques are applied to improve the relationship between professor and student. Ac-
cordingly, typologies of students will be obtained based not only on their navigation
and demographics but also on behaviour evolution.
Thus, five steps will be needed:
1. Data pre-processing: clean and prepare the Apache log file and load the clean data
into a relational database
2. OLAP analysis: design a multidimensional structure in which the main factors un-
der analysis: sessions, pages, user data, user behaviour (contents and navegation)
will be taken as dimensions and later build OLAP cubes in order to analyse the re-
corded data
3. Pattern discovery: application of data mining algorithms. Firstly descriptive algo-
rithms (clustering and association) will be applied to obtain typologies of users.
Then classification techniques will be used in a later step to classify behaviours ac-
cording to historical patterns
4. Pattern evaluation: All the patterns obtained will be evaluated to distinguish the
patterns that really help to better achieve the site goals.
5. Recommendation engine: integrate the discovered patterns in the online component
to provide personalized environment to learners.
Consequently, results of the project will make it possible to design improved e-
learning tools mainly in two aspects: on the educator side, their tasks will be devel-
oped in an informed way, ie, he will have measures to better follow the learning proc-
ess and to analyze the web course structure effectiveness; on the students side, they
will obtain a personalized environment that in a next future, will recommend them ac-
tivities and/or resources that would favour and improve their learning. An added value
of these tools will be that the site administrator will have parameters to improve the
site efficiency and adapt it to the behaviour of its users.
The development of this project will highlight important results, both methodo-
logical and application. A case study to show the obtained results will be presented.
The case study will be performed making use of the data from Universidad de Can-
tabria, and WebCT tool will be used. The project will be performed following the
CRISP-DM [1] process model. As data mining software mainly WEKA [7] will be
used, while Microsoft SQL Server will be used as the data repository.
References
1. CRISP-DM. http://www.crisp-dm.org
2. Dai, H., Mobasher, B. A road map to more effective web personalization: Integrating
domain knowledge with web usage mining. In Proc.of the International Conference on
Internet Computing 2003 (IC’03), Las Vegas, Nevada, June 2003.
3. Groeneboer, C, Stockley, D. Virtual-u: A collaborative model for online learning environ-
ments. 2nd Int Conference on Computer Support for Collaborative Learning, Toronto, Can-
ada, December 1997.
4. Hanna, M. Data Mining in the e-learning domain. Campus Wide Information Systems. 16
Jaunary 2004. vol. 21 no. 1 pp. 29-34(6). Research article ISSN: 1065-0741
5. Perkowitz, M., Etzioni, O. Adaptive web sites: Automatically synthesizing web
pages. In AAAI/IAAI, pages 727–732, 1998.
6. WebCT, http://www.webct.com/
7. WEKA http://www.cs.waikato.ac.nz/ml/weka/
8. Zaïane, O. Web Usage Mining for a Better Web-Based Learning Environment. Proc. of Con-
ference on Advantage Technology for Education. Alberta, Canada. 2001.
Similarity Queries in Data Bases
Using Metric Distances -
From Modeling Semantics to its Maintenance
Josef Küng1, Roland Wagner1

1 Johannes Kepler University of Linz
Institute for Applied Knowledge Processing (FAW),
Altenbergerstraße 69
4040 Linz, Austria
{jkueng, rrwagner}@faw.uni-linz.ac.at
http://www.faw.uni-linz.ac.at
Extended Abstract
Similarity search has emerged and become a fundamental requirement, even for data
base applications and data base management systems. A theoretical framework for
similarity queries is given in [1]. Unfortunately similarity search based on the attrib-
ute values itself does not fit all user needs. For example in the case of a tourism in-
formation system the attribute values “San Augustin” and “Playa del Ingles” should
be identified as similar because both are villages in the southern part of Gran Canaria,
even if their attribute values are anything but similar. Additional concepts to model
and maintain semantic background information are needed to support such enhanced
similarity queries.
There are several approaches to represent similarities which can’t be derived di-
rectly from attribute values. One of them is to use points in multidimensional spaces
as substitutes. Each attribute domain can have its own space. Attribute values are
represented as points in that space and the distances correspond reciprocally to their
similarity. These attribute similarities build the basis for computing whole objects
similarities.
At the institute for applied knowledge processing a prototype for such a system has
been implemented and tested. We called it VQS – Vague Query System [2]. One of
the main goals of this system was that it can be used as an add-on to still existing data
bases. Therefore we concentrated on relational and object relational data bases. The
semantic background information (the mapping from attribute values to multidimen-
sional points) lies in Numeric Coordinate Representation Tables (NCR-Tables).
These build quasi the similarity base, comparable to a knowledge base. An extra
query language (VQL – vague query language) allows formulating similarity queries.
It looks like SQL but, supports a new operator “is” which means “is similar to”. Any
combination of similarity conditions with non similarity conditions and, even similar-
ity joins are supported by this system. A practical test using the database of a large
tourism information system showed that the system works quite well.
One crucial question is left: “How to build an adequate representation of attribute
values in form of multidimensional points?” In some cases this mapping is obvious.
For localities we can use the geographical coordinates, for colors a three dimensional
representation (red, blue, green) could be chosen, and so on. For all other cases,
where such a simple mapping can not be found the NCR-table has to be maintained
manually. Newer research at our institute showed that an automatic finding of an
appropriate NCR-representation is just possible when all similarity values between
each pair of attribute values are known [3]. Otherwise – which is the case in practice
– for the number of dimensions needed just an upper bound can be given and the
finding of multidimensional points can be modeled as an optimization problem.
Therefore we are now working on using learning techniques such as reinforcement
learning for adjusting the similarity base step by step.
References
1. Jagadish, H. V., Menelzon A. O., Milo T.: Similarity-based Queries. Proceedings of the
fourteenth ACM SIGACT-SIGMOD-SIGART symposium on Principles of database sys-
tems, ACM, 1995
2. Küng, J., Palkoska, J.: VQS – A Vague Query System Prototype. Proc. Eighth International
Workshop on Database and Expert Systems Applications, September 1-2, 1997, Toulouse,
France, IEEE Computer Society Press (1997) 614–618
3. Csaji, B.C., Küng, J., Palkoska J., Wagner R.: On the Automation of Similarity Information
Maintenance in Flexible Query Answering Systems. Proceedings of the 15th International
Conference on Database and Expert Systems Applications - DEXA 2004, Springer, Lecture
Notes in Computer Science (2004) 130-140
A WEB-CASE Tool Prototype For Hybrid
Software Development
F. J. Orellana1, F. Guil1 , I. M. del Aguila1 , and S. Túnez1

Departamento de Lenguajes y Computación
Universidad de Almería
04120 Almería
{fjorella, fguil, imaguila, stunez}@ual.es
Abstract. In this paper we present a Web-based CASE tool for hy-

brid software projects that supports rst stages of InSCo, an extended
methodology based on CommonKADS and RUP. The tool InSCo Requi-
site will guide the development of a sort of software where knowledge-
based components are integrated with traditional information systems.
Furthermore, this tool will allow to manage several development projects
and select the users that will take part of them, making easier the coo-
perative work.
1 Introduction
When we are developing software, based or not based on knowledge, develo-
pers choose a methodology depending of the problem, the prior knowledge, the
available tools or other factors. These methodologies lead us to the production
process determining the documents and artifacts we must generate, the activities
to do and the order that we must follow.
The main artifact generated during the software development process, is a
group of models [1]. The models of rst stages in the process are often specied
using natural language. However, this technique has problems like ambiguity, un-
certainly or imprecision. Several alternatives have proposed a solution by means
of the utilization of a structured natural language, description language, des-
cription languages of design, graphic notation or formal specications, that are
similar to the successfully used in the nal stages of software development [4].
We propose to use these techniques in the organization modelling and requi-
site level scope, combining the forms and diagrams to model complex systems
and a CASE tool for managing these artifacts. But, in a lot of cases, the models
and documentation that we generate and maintain is very complex. So, we need
software tools that guide us in the generation of these artifacts, its maintenance,
and the control of the integrity of changes.
2 The InSCo Requisite Tool

In many application domains, there are problems where we need to build a
software system that uses knowledge engineering methods. Usually, the built
system requires some extra functionality that must be resolved using a tradi-
tional information system [2]. For this reason, a methodology has been proposed
integrating both techniques, knowledge engineering and software engineering, to
reach the best results [3]. The InSCo methodology was designed from a requisite
engineering perspective to deal with the hybrid software development. This is
a kind of software where knowledge-based components and information systems
are combined in a unique solution. As we can see in Figure 1, InSCo proposes
a total of seven main activities for developing hybrid software: Organization,
Requisite, Knowledge, Analysis, Design, Implementation and Maintenance or
Evolution. The prototype we have developed covers the two rst: Organization
and Requisite. The results of these two activities are the business and services
models. The business model covers the organizational environment, the roles
that take part in the project, the objectives of the organization, the information
resources and nally, a list of items for making decisions about the viability of
the project. On the other hand, the services model represents a more detailed
specication of the information that we can obtain from the business model. This
specication is related to functional, non-functional and information requisites
of the project.
Fig. 1. Models of InSCo Methodology
A competitive methodology must have a set of associated tools for making

easier the development of software based on it. In this paper we present the
InSCo Requisite tool, designed for helping us in the form constructions, and the
administration of diagrams associated with the modelling of our hybrid software
development methodology. It supports the control of several projects which are
managed by a group of users via the Internet.
Mainly, there are three kind of users that can interact with InSCo Requisite.
First of them, the system administrator, whose main activity is the management
of the projects of hybrid software development, adding new projects, modifying
their information or removing them. Furthermore, this kind of user is in charge
of establish an administrator for each project. Next, we can nd the project
administrator role, who is the person that leads a development project and
decide who are the users that are going to take part in the project. Besides,
this person can participate in the project like a developer. Finally, the third
class are the developers that take part in the dierent projects. Their main
activity is to work with the whole group of forms that belong to the rst levels
of InSCO methodology. They are in charge of managing the documents and
artifacts associated with each project where they belong.
Due to InSCo Requisite manages several projects, that can be formed by a
group of users and possibly placed in dierent areas of the world, we selected a
Web-based architecture, using JSP and JavaBeans technology for its implemen-
tation, being Oracle the DBMS.
Figure 2 shows a screenshot of InSCo Requisite. On the left we can see a
tree-based menu, which represents the hierarchy of the forms that belong to a
specied project. On the right side, we can see a sample of a form of InSCo
methodology.
Fig. 2. Screenshot of InSCo Requisite
References
1. Object management group. In MDA Guide Version 1.0.1. OMG Document:
omg/2003-06-01, 2003.
2. J. J. Cañadas, A. Bosch I. M. del Águila, and S. Túnez. An intelligent system
for therapy control in a distributed organization. In Proceedings of Information and
Communication Technology (EurAsia-ICT 2002), Springer-Verlag. Lecture Notes in
Computer Science, 2510:1926, 2002.
3. I. M. del Águila, J. J. Cañadas, J. T. Palma, and S. Túnez. Integration of develop-
ment methodologies for the building of knowledge intensive multiagent systems. In
In Proc. of the First Int. Conf. on Integration of Knowledge Intensive Multi-Agents
Systems, KIMAS'03, pages 203208. IEEE Press, 2003.
4. G. Schreiber, H. Akkermans, A. Anjewierden, R. deHoog, N. Shadbolt, W. Van-
deVelde, and B. Wielinga. Knowledge Engineering and Management: The Com-
monKADS Methodology. The MIT Press, 1999.
An Augmentative Communication System based on
Adaptive Evolutionary Hypermedia Systems
M.V. Hurtado1, N. Medina1, L. García2, , M.L. Rodríguez1

1
Dept. Lenguajes y Sistemas Informáticos. University of Granada (SPAIN)
1
Dept. Informática. University of Jaén (SPAIN)
1
{mhurtado, nmedina, mlra}@ugr.es} 2lina@ujaen.es
1 Introduction
Augmentive Communication Systems [1] (sign languages, pictorial languages, sign

templates and communicators) are part of a technology developed to improve the
social integration of people with temporal o permanent communicative difficulties
providing useful tools for their rehabilitation.
Each person requires a specific attention and adaptation to fit his profile, which
could evolve in time. However these systems don’t support mechanisms for their
necessary adaptation to each user and situation taking in account his capabilities,
skills and progresses at run-time. Therefore, an efficient and suitable approach to
design these systems is needed, based on software evolution and adaptability [2,3].
In particular this paper proposes an adaptive evolutionary hypermedia (from a
knowledge and an interaction models) that allows to represent control and to adapt the
context communication.
2 Features of Adaptive Evolutionary Communication Systems
The elements that this approach manages in order to design adaptive evolutionary
communication systems are:
• Ontology model, which represents each conceptual world and describes a
specific knowledge domain by means of concepts and relationships between them
(a semantic net) [4]. The ontology, in this case, provides the link between the
symbols used in the communicator (image, sound or text adapted to each
person’s knowledge) and the entities of the real world they represent. In addition,
ontologies both allow a global vision of each conceptual world based on the
concepts employed by the users, and allow to standardize into a single model the
knowledge about the people with the communicative difficulties, the knowledge
about the people around them (i.e. relatives, tutors, rehabilitators, etc.), and the
media used for communication (i.e. templates, agendas, etc.).
• User model, which provides knowledge about the people with communication
disabilities and consists on the following components:
o Knowledge domain represented by means of an ontology. In addition, the
educators can define different semantic views of the whole knowledge
domain taking into account the user profile. For example, figure 1 shows a
partial view of an ontology that describes a scenario of shopping.
Fig. 1. Shopping scenario
The process followed by the tutor to create a new view consists in hiding
in the semantic net the concepts, items and conceptual relations he
considers not related to the current view. The system automatically keeps
the integrity of the view, for example, if after hiding a relation some
concepts get disconnected, they are also automatically hidden. In the same
way, when a concept is hidden all the items associated to it and all the
relations starting or arriving to the concept are also hidden.
Once the semantic net is built, the system will be able to automatically
generate the interaction templates from it. These templates constitute one
of the possible media used to establish the communication with the user
and are represented in hypermedia format.
o User profile, containing the particular characteristics of users
(communication habits, training or therapy objectives, and interaction
preferences). To construct and select it the user (and relatives or educators)
must provide this information. This artefact allows the adaptation and
selection of the semantic view that best fits to a particular user.
• Adaptation by feedback. By means of a transition matrix, the system captures
the conceptual relationships followed by the person during navigation. The
matrix has one row and one column for every concept represented in the
knowledge domain. Based on the analysis of the matrix, the system can identify
new conceptual relationships that are needed or existing conceptual relations that
should be removed in the knowledge domain. In addition, this mechanism helps
to detect progress or problems in the evolution of the patients.
3 Architecture of Adaptive Evolutionary Communication System
The architecture of our adaptive model has two tiers: the system (communicator) and
the meta-system [6]. This division allows us to separate on the one hand the
interaction, communication and user adaptation and on the other hand its evolution
process.
The communicator is used by the user and the meta-system by the educators. The
cognitive, interaction, design and adaptation aspects have to be differentiated to avoid
the coupling. Thereby, evolution/adaptability can be done more easily and safely
because the architecture components are independent.
The system is itself divided in four parts, which allows us to tackle separately the
aspects of knowledge representation (memorization system), construction of partial
views focused to one o several concrete knowledge subdomains (presentation
system), navigation of the built views (navigation system), and personalization of the
navigation process (learning system) [2].
The structure offered to navigate is different depending on the used platform. If a
personal computer is used the interface has two frames: at the left the semantic net
and at the right some information about the element selected in the net. It provides a
semantic and contextual mode of navigation. If it has not been predefined by the
educator, the user can choose the navigation mode that he desires. Depending of the
chosen mode (free navigation, conceptual navigation, navigation by conceptual
relation or navigation by prerequisites), some options in the semantic net are disabled
according to the user features and his interests. So, the adaptation is carried out
depending on the user, but also on the navigation mode used in that moment.
In other case, when a PDA or other embedded system is used, the interface is based
in templates. Every template only shows the available options in the current moment,
without additional information. Each option is represented by means of an icon which
can also have associated text and audio.
Figure 2 shows a possible use of the PDA in a scenario based on figure 1. As it can
be seen, in some templates all the options must be selected in a particular sequence,
while in others it is enough that the user selects one or several options in any order.
Fig. 2. Shopping with the PDA
The meta-system is in charge of the evolution of the system. It includes the

evolution of the complete knowledge domain, of the partial views built from it, and of
the navigation and adaptation rules used during the user interaction. For example,
when a concept is removed in the knowledge domain, the meta-system removes this
concept in all the views that show it. In addition, if after the change, some concept is
disconnected in a semantic net, the meta-system performs new modifications in order
to guarantee the consistency of the net.
References
1. Schlosser, R.W. Braun, U. Efficacy of AAC Interventions: Methodological Issues in

Evaluating Behaviour Change, Generalization and Effects. ACC Augmentative and
Alternative Communication, 10, 207-233. 1994.
2. Medina, N. Molina, F. Rodríguez, M.J. García, L. An Architecture for the Development,
Evolution and Adaptation of Hypermedia Systems. Proceedings of the International
Conference on Software Engineering Research and Practice. SERP'03. Pp: 112-118. Las
Vegas, Nevada, USA. June 23-26, 2003.
3. Rodríguez, M.J. Paderewski, P. Rodríguez, M..L. Gea, M. An approach for evolving an
adaptive and unanticipated system: a communicator for children with autism. FUSE 2004.
Workshop on Foundations of Unanticiped Software Evolution, pp. 40-47, 2004.
4. Hurtado, M.V. García, L Parets, J. Semantic Views over Heterogeneous and Distributed
Data Repositories: Integration of Information System based on Ontologies. Advantages In
Knowledge Organization. Vol. 8, pp 332-347. Ergon Verlag , 2002.
5. Rodríguez, M.L. Rodríguez, M.J. Gea, M. A Framework for Modelling the User Interaction
with a Complex System. Computer Aided Systems Theory-EUROCAST 2003 LNCS 2809.
pp: 50-61. Springer Verlag, 2003.
6. García, L. Rodríguez, M.J. Parets, J. Evolving Hypermedia Systems: a Layered Software
Architecture. Journal of Software Maintenance and Evolution: Research and Practice.
Wiley. Vol 14. pp: 389-406. 2002.
A Tool for Thematic Cartography of Corpora
Thibault Roy and Pierre Beust
GREYC, Team ISLanD, Computer Science Laboratory, University of Caen, F14032, Caen
Cedex, France
{troy, beust}@info.unicaen.fr
1 Introduction
In our researches in Computer Science and particularly in Natural Language

Processing (NLP), we are interested in electronic management of textual documents.
For a lot of tasks of extraction and retrieval of information, the discovery of themes in
sets of documents is an important and difficult analysis. The ProxiDocs1 tool [1] helps
its users in such tasks allowing them to visualize graphical representations (called
here “thematic maps”) of documents’ sets in taking account of themes chosen and
specified by users. These maps allow users to discover thematic differences and
similarities existing between each document of the analyzed set. By the term "themes"
we consider the main subjects dealing in a corpus of documents [2].
In the first section, we’ll give an overview of tools using visualization methods in
order to access information of documents’ sets. The second section will deal with the
mechanism of ProxiDocs and thematic maps returned by the tool.
2 Textual data visualization
The quantity of information exchanged every day on the Web doesn’t stop increasing.
Generalist search engines, like Google or Yahoo, return hundreds of links from a
user’s request. These links are ordered in lists, but such lists do not allow users to
have a global view on returned results. However, such view would reveal the
principal categories of information from returned Web pages. In order to give users a
global view on documents’ sets, it can be interesting to use graphical representations
[3]. A lot of NLP tools represent sets of documents in 2 or 3-Dimensional space (we
can cite [4], [5], [6], [7], for such works), each one uses specific visualization
methods.
The visualization method used in ProxiDocs is calling “cartography”. A map of a
set of documents reveals proximities and links between textual entities (like words,
documents, etc.); we can compare this visualization method with road maps which
reveals proximities and ways between towns. Since 2001, the two metasearch engines
KartOO [8] and MapStan [9] are available on the Web2. These tools return a map
1 ProxiDocs is available at: http://www.info.unicaen.fr/~troy/proxidocs.

2 KartOO and MapStan are respectively available at: http://www.kartoo.com and
http://www.mapstan.net.
from the user's request; this map represents Web pages related to the request. For
these systems, similar pages are located in a same zone on the map. Also, it's possible
to distinguish principal categories of information related to the user’s request.
3 ProxiDocs, an NLP tool for thematic cartography
ProxiDocs returns thematic maps from a set of documents (for example: a corpus of
newspaper's articles or a set of Web pages returned by a search engine) and themes
specified by the user. A theme will be defined by a set of words chosen by the user.
For example, the theme “Sport” could be specified by the following terms: “winner”,
“looser”, “tennis”, “football”, “Zinedine Zidane”, etc. We advise the ThemeEditor3
tool [10] develop in our research team for a such task of themes contruction’s.
The first stage consists in counting words of each theme in each document. So, we
associate a numbers’ list with each document, and, each list constitutes a N-
Dimensional vector (N is the number of themes specified by the user). For instance, if
a document contains 3 words of the theme “Economy”, 2 words of the theme “Sport”
and 7 words of the theme “Education”, the vector associated to this document is (3, 2,
7).
The next stage consists to project these N-Dimensional vectors on a 2-Dimensional
space. In order to realize a such projection, we use a method called Principal
Components Analysis [11]. Each document is then represented by a point in a 2-
Dimensional space, so, first maps can be constructed. On these maps, proximities
between different points could inform the user on thematic similarities between
different documents.
In order to underline documents’ subsets with similar themes, we use a clustering
method called Ascendant Hierarchical Clustering [11]. Thus, thematic maps reveal
clusters of documents, each cluster containing documents dealing with similar
subjects.
An experimentation of ProxiDocs from a corpus of around 800 articles of the French

newspaper “Le Monde” of 1989 (around 700,000 words) and a generalist set of
themes can reveal the two following kinds of maps:
3 ThemeEditor is available at: http://www.info.unicaen.fr/~beust/ThemeEd/.

Fig. 1. Each point on the left map represent a document of the analyzed corpus, each disc on
the right map represent a subset of documents. The height of a disc is proportional to the
number of documents included in the subset. A color related to the principal theme of
documents or sets of documents is attributed to each point and disc. Each point is an hyperlink
to the represented document, each disc is an hyperlink to a document outlining the themes of
documents of the represented set.
4 Conclusion
Proxidocs allows its users to visually identify main thematic tendencies of a set of
documents. We work now on the application of thematic cartography to results of
search engines in order to distinguish principal thematic categories of information
contain in returned Web pages. In an technological watch context, our tool could also
reveal thematic changes of a set of documents regarding to the time. In order to
improve our tool, we have to measure the “quality” of thematic maps returned to
users. This assessment will be difficult because we will have to elaborate a protocol
taking account of the point of view of each user on themes used during the
construction of thematic maps.
References
1. Roy T., Beust P., ProxiDocs, un outil de cartographie et de catégorisation thématique de

corpus, Proceedings of the 7th International Conference on the Statistical Analysis of
Textual Data, pp. 978-987, 2004.
2. Pichon R., Sébillot P., Différencier les sens des mots à l'aide du thème et du contexte de
leurs occurrences: une expérience, Proceedings of Natural Language Processings, pp. 279-
299, 1999.
3. Shneiderman B., The eyes have it: a task by data type taxonomy for information
visualization, Proceedings of Visual Languages, Boulder, 1996.
4. Salton G., Automatic Text Processing: The Transformation, Analysis, and Retrieval of
Information by Computer, Addison-Wesley, Reading, Pennsylvania, 1989.
5. Hearst M. A., Tilebars: Visualization of term distribution information in full text
information access}, Proceedings of ACM SIGCHI, pp. 59-66, 1995.
6. Robertson G.G, Mackinlay Jock D. and Stuart K.C., Cone Trees: Animated 3D
Visualizations of Hierarchical Information, Proceedings of ACM SIGCHI, pp.189-194,
1991.
7. Lamping J., A focus+context technique based on hyperbolic geometry for viewing large
hierarchies, Proceedings of ACM SIGCH, 1995.
8. Chung W., Chen H., Nunamaker J.F.Jr.: Business Intelligence Explorer : A Knowledge Map
Framework for Discovering Business Intelligence on the Web, Proceedings of the 36th
Hawaii International Conference on System Sciences , 2002.
9. Spinat E., Pourquoi intégrer des outils de cartographie au sein des systèmes d’information
de l’entreprise ?, Colloquium of Information Cartography, 2002.
10.Beust P., Un outil de coloriage de corpus pour la représentation de thèmes, Proceedings of
the 6th International Conference on the Statistical Analysis of Textual Data, pp. 161-172,
2002.
11.Bouroche J.M., Saporta G., L’analyse des données, Collection Que sais-je ?, PUF, 1980.
System-oriented Evaluation for Multi-tier
Personalization of Web Contents *
Alberto Díaz1, Pablo Gervás2 y Antonio García3

1CES Felipe II – Universidad Complutense de Madrid
adiaz@cesfelipesegundo.com
2
SIP – Universidad Complutense de Madrid
pgervas@sip.ucm.es
2
Departamento de Comunicación - Universidad Rey Juan Carlos
an.garcia@cct.urjc.es
1 Introduction
A Web contents personalization system is based on 3 main functionalities: content

selection, user model adaptation, and content generation. These functionalities must
be based on a user model. Content selection refers to the choice of the particular sub-
set of all available documents that will be more relevant for a user model. User model
adaptation is necessary because user needs change over time as a result of his interac-
tion with information [1]. In our case, content generation involves generating a new
Web document that contains, for each selected document, its title, its relevance, a
personalised summary, and a link to the full document.
In this paper we focus on the evaluation of the selection and adaptation processes.
We perform a system-centred evaluation based in the precision and recall metrics
obtained through different configurations of the system. The aim is to identify which
is the best way of carrying both processes: selection and adaptation.
2 The multi-tier proposal
The multi-tier content selection process to be employed in this paper involves a do-
main specific characterization, a set of keywords, a relevance feedback tier, and an
automatic categorization algorithm.
The first tier of selection corresponds to a domain specific given classification of
the documents - for instance, in a digital newspaper, the assignment of news items to
newspaper sections: national, international, sport, etc. The user must assign a weight
to them (su).
For the second tier, the user enters a set of keywords - with an associated weight -
to characterise his preferences: appearance of these keywords in the documents indi-
* This research has been partially funded by the Spanish Ministerio de Ciencia y Tecnología
(TIC2002-01961)
cates a possible interest. These keywords are stored, for each user u, as a term weight
vector (ku).
For the third tier the user must choose - and assign a weight to them - a subset of
those categories in the first level of Yahoo! [4]. This information is stored as a matrix
where rows correspond to categories and columns correspond to users (Ggu). These
categories are represented as term weight vectors (g) by training from the very brief
descriptions of the 14 first level of Yahoo! Spain categories entries [4].
In the fourth tier, short term interests [1, 5] are represented by means of feedback
terms obtained from user provided feedback over the documents he receives [2]. The
term weight vector for each user (tu) is taken to represent the short term interests of
that user. These correspond to information needs whose interest to the user wanes
after a short period of time, so their weight must be progressively decreased over
time.
Documents are downloaded from the web as HTML documents. For each docu-
ment, title, category, URL and text are extracted and stored for subsequent process-
ing. Term weight vector representations (dd) are obtained by application of stop lists,
stemmer, and the tf · idf formula for computing actual weights [6].
3 Selection and Adaptation
As each web document has a preassigned section, it is assigned the weight associated
with the corresponding section in the particular user model. The similarity between a
document and a category is computed using the cosine formula of the vector space
model [6, 7]. The similarity between a document and the keywords of a user model is
computed also using the cosine formula.
⎛ 14 ⎞
⎜ ∑ Giu sim(d d , gi ) ⎟
α su + β ⎜ i =1 14 ⎟ + χ sim(d d , ku ) + δ sim(d d , tu )
⎜
⎜
⎝
∑
i =1
Giu ⎟⎟
⎠
(1)
rdu =
α + β + χ +δ
The short term model is obtained as a result of the process of adaptation of the user
model. The user receives a web document that contains an automatically generated
personalised summary [3] for each of the 10 web documents that the system has
found more relevant according to his user profile. With respect to this information the
user may give positive or negative feedback. The feedback terms of the short term
model are obtained from the news items for which feedback has been provided [2].
The similarity between a document and the feedback keywords of a user model is
computed also using the cosine formula. When all documents have been ordered with
respect to the various reference frameworks, the results are integrated using a particu-
lar combination of them.
4 Evaluation
Experiments are evaluated over data collected for 106 users and the news items corre-
sponding to three weeks – the 14 working days of the period 1st-19th Dec 2003 - of
the digital edition of the ABC Spanish newspaper. The average of news items per day
is 78.5.
To carry out the evaluation, judgements from the user are required as to which
news items are relevant or not for each of the days of the experiment. To obtain these
judgements users were requested to check the complete set of news items for each
day, stating for each one whether it was considered interesting (positive feedback),
not interesting (negative feedback) or indifferent (no feedback). Users were explicitly
asked not to confine their judgements on interest to relevance with respect to the
initial user profiles they had constructed on first accessing the system, but rather to
include any news items that they found interesting on discovery, regardless of their
similarity with respect to their initial description of their interest. It is hoped that
enough information to cover these rogue items will be captured automatically and
progressively by the system through the feedback adaptation process.
Since our experimental set up combines a binary relevance judgement (indifferent
is considered as not interesting ) from the users and a ranking of news items provided
by the system, it was decided to use normalised recall and precision [6] as our evalua-
tion metrics. In addition, with respect to equal relevance values for consecutive posi-
tions of the ranking, the average ranking of the whole set of conflicting positions has
been taken as ranking for each and all of them. This adjustment avoids the problem of
ordering items at random within the ranking when they have equal relevance.
The following experiments have been carried out to check the validity of the pro-
posed model for the selection process. Each experiment combines different possibili-
ties for long term modeling - only sections, L(S), only categories, L(C), sections and
categories, L(SC), sections and keywords, L(SK), categories and keywords, L(CK),
sections, categories and keywords, L(SCK). This implies giving different values to
the parameters α, β, χ and δ of formula (1). For example, L(SC) is the combination of
sections and categories for the long term model (α=1, β=1, χ=0, δ=0).
Table 1. Global averages (per user and per day) and relative increments in normalised precison
and recall for different configurations for the long term model for the selection process
L(SCK) L(SC) L(CK) L(C) L(SK) L(S) L(K)

Pr 0.588 0.579 0.515 0.505 0.497 0.450 0.351
% Pr 1.7 14.3 16.4 18.3 30.9 67.8
R 0.685 0.680 0.617 0.605 0.649 0.637 0.531
%R 0.8 11.1 13.2 5.5 7.6 28.9
The results shows (Table 1) that, in the case of the selection, the combination of
sections and categories with keywords is better, in normalised recall and precision,
than whatever other combination.
For the validation of the adaptation process, some experiment have been carried
out combining different possibilities for long term modeling, either acting on their
own or in combination with the short term model, S. For example, L(SC)S is the
combination of long and short term model when the long term model includes sec-
tions and categories (α=1, β=1, χ=0, δ=1).
Table 2. Global averages (per user and per day) and relative increments in normalised
precision and recall for different configurations for the long term model for the adaptation
process
L(SCK)S L(SC)S L(SK)S L(CK)S L(C)S L(S)S L(K)S S

Pr 0.600 0.583 0.568 0.539 0.535 0.514 0.475 0.421
% Pr 2.9 5.6 11.3 12.3 16.7 26.3 42.4
R 0.691 0.681 0.669 0.633 0.614 0.652 0.583 0.545
%R 1.5 3.2 9.1 12.6 6.0 18.4 26.9
It can be observed (Table 2) that the combination of long and short term is better
than if they are used alone, in terms of normalised recall and precision, and that the
best results in the combination of the short and long term models are obtained when
sections, categories and keywords are combined for the long term model.
References
1. Billsus, D. & Pazzani. M.J., 2000. “User Modeling for Adaptive News Access”, User
Modeling and User-Adapted Interaction Journal 10(2-3), pp. 147-180.
2. Díaz, A. & Gervás, P., 2004. “Adaptive User Modeling for Personalization of Web Con-
tents”. Proceedings of AH 2004, LNCS 3137, pp. 65-75.
3. Díaz, A. & Gervás, P., 2004. “Item Summarization in Personalization of News Delivery
Systems”. Text, Speech and Dialogue, TSD 2004, LNAI 3206, pp. 49-56.
4. Labrou, Y. & Finin, T., 2000. “Yahoo! As an Ontology: Using Yahoo! Categories to
Describe Documents”. Proceedings of CIKM-99, pp. 180-187, ACM Press.
5. Mizarro, S. & Tasso, C., 2002. “Ephemeral and Persistent Personalization in Adaptive
Information Access to Scholarly Publications on the Web”. 2nd International Conference
on Adaptive Hypermedia and Adaptive Web Based Systems, Málaga, España, 2002.
6. Salton, G., 1989. Automatic Text Processing: The Transformation, Analysis and Retrieval
of Information by Computer, Addison-Wesley Publishing, Reading, Massachussets, 1989.
7. Sebastiani, 2002. “Machine learning in automated text categorization”.
ACM Computing Surveys (CSUR), 34(1), pp. 1-47, ACM Press, 2002.
The gaps of the thesaurus WordNet used in
Information Retrieval
Javier de la Mata, Jose A. Olivas, Jesus Serrano-Guerrero
Computer Science Dept.

University of Castilla-La Mancha.
Paseo de la Universidad 4, 13071
Ciudad Real, Spain.
JavierL.Mata@alu.uclm.es, Joseangel.olivas@uclm.es,
Jesus.Serrano3@alu.uclm.es
1 Introduction
Current Internet search engines try to solve the main problems that affect the quality of
the obtained results for getting a relevant collection of documents. Different users may
dis agree about the relevance of the obtained documents using the same query.
Relevance seems to be a subjective concept, mainly due to the ambiguity and the vo-
cabulary problems. Then, the search engine needs to know some kind of semantic
information that allows it to improve the results.
The main sources that are used to discover the semantic information are the dic-
tionaries, thesaurus or ontologies. WordNet [1] is one of the main tools used in infor-
mation retrieval processes, mainly for the meaning disambiguation tasks. This tool have
several problems such as, for example, the problem of the granularity of the senses or
the lack of recent terms. For example, WordNet 2.0 does not recognize the terms “XML”
or “CORBA”. Other times, it recognizes the common senses of one word, like in the
case of “SOAP”, but not recognizes a new sense (Simple Object Access Protocol).
This work is focused on the study of the weak points of WordNet. Our objective is
to use this information to improve our knowledge structures. For this reason, the study
of the user’s query is an important aspect that helps us to know the terms not recog-
nized by WordNet. This information can be later on used for getting better search
results.
2 GUMSe Overview
Meta-search engines offer a unified way to access multiple search engines, avoiding
the user to interact with all the underlying engines [2]. Taking advantage of the fea-
tures that they provide, we have developed GUMSe like a platform that allows us to
test new ideas applicable to the Web search process. The GUMSe architecture (Fig. 1)
is composed of a set of agents that communicate to each other by means of a network.
The main agents that form the architecture are Search Broker Agent, Sense Resolver
Agent, Query Agent Expansion, Rater Web Page Agent, Search Engine Agents, Con-
struction Ontology Agent and WordNet Agent.
Fig. 1. GUMSE architecture.
Using the classic technique of query expansion, we semantically obtain a collection

of additional queries, constructed by the Query Expansion Agent, related to the origi-
nal one. New queries are generated, replacing or introducing new related terms to pre-
vious ones by means of synonymy, hyponymy or hyperonymy relations. It is neces-
sary to notice that it is difficult to obtain the semantic related concepts without the user
collaboration in queries with only one term. This situation is solved by assuming the
most commonly used sense or asking the user for suitable senses. In the queries with
several terms we try to obtain the underlying semantic relations using WordNet,
without user intervention if it is possible.
3 Motivation
The aim of this work is to discover the user tendencies when they use the search en-
gines. With this information it is possible to design a more suitable evaluation algo-
rithm of the retrieved web pages. For this reason, we study the set of keywords and
queries more frequently used in the search engines.
There are many works focused on the user search techniques. Most of these works
show statistics about the number of keywords used, the number of queries per session,
and other statistical data [3]. Our objective is to discover the knowledge gaps of
WordNet. This information could be very useful to avoid negative situations in our
system. The main usual disadvantages are two: errors in the disambiguation process
and lack of information about the query terms.
Nowadays, the users often use the search engines to find the web page about a
concrete product, artist or program. In these cases, the traditional knowledge sources
fail because this information is about a specific domain, such as music or Internet. For
this reason, the construction of new knowledge structures that help the search engine
to know the specific domain of the user’s queries could be an important task. If the
search engine knows the domain, it is possible to modify some characteristics of the
search process. For example, if one user looks for one musical band and the system
knows it, the system could assign more importance to the results of one specialized in
music search engine. Another possibility is to introduce generic terms about the do-
main in the query expansion that avoid retrieved irrelevant web pages.
Usually, search engines offer tools for knowing the more consulted topics. Also,
there are many search engines and web pages that show the top searches. The data
used in our study are from a variety of sources. The main two are HitBrain1 and
Search.com2, among others.
4 Conclusions
In our preliminary study, WordNet recognizes the 45,4% of the more frequent used in
the user queries keywords. In this study, we used 10.000 keywords obtained from
HitBrain. The keywords have between 1 and 5 words. Another important issue is that
there are situations where WordNet recognizes keywords, but in a wrong way. These
cases are very infrequent but they produce mistakes in the search process.
It seems necessary to have good knowledge structures that can be used in informa-
tion retrieval to focus the search for obtaining good results. For this reason, the here
proposed subsystem, specialized in the improvement of the knowledge structures
could be a good support for Internet searchs.
Acknowledgements
Partially supported by PBC-03-004 PREDACOM project, JCCM, TIC2003-08807-C02-02
DIMOCLUST project, MCYT, and JCCM Cons. de Ciencia y Tecnología - European
Social Fund grants, Spain.
References
1. Miller, G.: Special issue, WordNet: An on-line lexical database, Intl. Journal of Lexicogra-
phy 3(4), 1990.
2. Gómez , M., Abasolo, J. M.: Improving meta-search by using query-weighting and nu-
merical aggregation operators. In Proc. of the Information Processing and Management of
Uncertainty conference, IPMU 2002, Annecy, France (2002): 1383-1390.
3. Jansen, B. J., Spink, A.: An Analysis of Web Documents Retrieved and Viewed. Interna-
tional Conference on Internet Computing (2003): 65-69.
1 http://www.hitbrain.com
2 http://www.search.com
Language for Uncertainty and Conflicting Agents
Germano Resconi1, Javier Alonso2,Raul Izquierdo2

1
Catholic University,via Trieste 17 Brescia,Italy
resconi@numerica.it
2
Object Oriented Technology Laboratori Department of Conputer Science (University of
Oviedo ) c/Calvo Sotero s/n 33005 Oviedo (Spain)
j.alonso@ceb2b2000.com,raul@a4devis.com,aajuan@uniovi.com
Extended Abstract
The traditional Object Oriented Technology is not sufficient agile at all to assume a
rapid development process. There is a lack of efficiency from the design to the
codification phase. This is due to the encapsulation of all the elements inside the
domain in one unique module: the class.
The RDM (Rapid Domain Modelling) architecture proposes another way for
transferring to code the different domain elements, following the Specialization
Principle and communication mechanisms through the Variable State Pattern. These
features are implemented in Entities, Relations and Operations, the main elements of
the RDM structure.
In this paper we want introduce a new dimensions inside the Object Oriented
language by the new language as the RDM. The new dimensions are uncertainty and
adaptation of the objects. Adaptation tries to reduce the uncertainty that came from
communication. Inside the OO we introduce new types of dynamic that balance the
uncertainty with the adaptation. Every agent perceives other agent’s behaviour as
coherent or incoherent in comparison with its particular behaviour (pattern). In the
communication among agents incoherence are possible. Incoherence means that for
the same object and the same property one agent give different and conflicting logic
valuation. A total coherence means that the property of the object is perceived from
the different agents from different points of view in the same way.
A set of True and False logic value for the same property give us the structure of
the uncertainty of the agent that can be formalise by special type of fuzzy logic
Adaptation or compensation try to change the structure of the uncertainty and also the
type of the fuzzy logic that agent use for the reasoning.
References
1. Izquierdo Castanedo, Raúl. “RDM: Arquitectura software para el modelado de dominios en

sistemas informáticos”. Doctoral Thesis. 2002 (Only available in Spanish) You can
download it in PDF version in the author homepage: http://di002.edv.uniovi.es/~ric/tesis.pdf
2. Kent Beck. “Smalltalk Best Practice Patterns”. Prentice Hall PTR;ISBN: 013476904X; 1st
edition (October 3, 1996)
3. G. Resconi, G.J. Klir and U. St. Clair, Hierarchical Uncertainty Metatheory Based Upon
Modal Logic, Int. J. of General Systems, Vol. 21, pp. 23-50, 1992.
4. G.Resconi Lakhmi C. Jain, Intelligent Agents, Speinger –Verlag 2004,Series in Fuzziness
and Soft Computing, vol.155
A Model-Based Architecture for Fuzzy Temporal
Diagnosis ?
J. M. Juarez, J. Palma, M. Campos, J. Salort, A. Morales, and R. Marin
Artificial Intelligence and Knowledge Engineering Group.

University of Murcia. Spain
jmjuarez@um.es
Extended Abstract
The design of knowledge-based systems has received special attention from the begin-
nings of Artificial Intelligence. In particular, the application of model-based systems has
obtained important results in the last decades [1]. However, the experience in develop-
ing these systems, far from becoming simpler, reveals the complexity on the design, the
development and its maintaining. Furthermore, in most of the cases, systems are highly
dependent on the domain and the problem to solve. The golden goal of a general model-
based architecture for all problem and domain is unreachable at the moment because of
the wide of the problem space to cover. However, general architectures fulfilling some
domain constraints could cover some parts of this problem space. Thus, in this work
we present a general model-based framework for intensive-knowledge domains, where
the temporal dimension (managing uncertainty) is considered to deal with the diagnosis
problem. The aim of this work is to present a general framework for temporal model-
based diagnosis. Nevertheless, the task of designing a general architecture is a complex
issue. One of the most difficult steps is to establish the set of goals to reach. Our general
framework tries to deal with the following key topics: domain independence, the man-
agement of temporal uncertainty, the knowledge acquisition bottleneck, model-based
representation, and the diagnosis task. In order to cover these points, firstly it is im-
portant to describe the main structure of the architecture, describing each part in detail
afterwards. This structure defines two modules, the Temporal/Domain Module and the
Decision Support Module.
The Temporal/Domain Module is composed by all those elements that permit a

domain independent architecture and the management of the temporal dimension. The
domain ontology server allows the rest of elements of the architecture to access the on-
tology of the domain. This ontology guarantees that all concepts used in the architecture
are semantically consistent. It is also important to mention those elements that allow the
temporal management. The temporal database stores all the system information as well
as some temporal tags. However, those tags are not enough to manage the temporal di-
mension. Thus, the use of temporal reasoning techniques provides, to the architecture,
?
This work was supported by the Spanish MCyT under project MEDICI (TIC2003-09400-C04-
01), the Murcia Regional Government under project PB/46/FS/02, and the Spanish MEC under
the FPU national plan (grant ref. AP2003-4476)
important advantages: to guarantee the temporal consistency, to infer new temporal re-
lations, and to carry out quite complex temporal queries. Hence, this architecture relies
on FuzzyTIME [2], a general purpose temporal reasoner to manage temporal uncer-
tainty and which is based on the formalism of the fuzzy temporal constraint network
(FTCN).
The Decision Support Module includes all elements needed to obtain the diagno-
sis solution. This architecture is proposed for a model-based diagnosis. It is, therefore,
needed a formal model, which is used and stored in the knowledge base of the archi-
tecture. This knowledge base is supported by some knowledge acquisition tools, which
must access both the formal model and the domain ontology. Those tools are in charge
of the storage of the expert’s knowledge, and intended for the smooth of the problem of
the knowledge acquisition bottleneck. The diagnosis is the key process of the module.
This task, described in the architecture, is a temporal model-based diagnosis. To this
end, the role of the diagnosis agent is to find a diagnosis solution from: the temporal
data (from the temporal database), the domain ontology (from the ontology server), the
knowledge base, and the help of the temporal reasoner. This process is based on an al-
gorithm that obtains a diagnosis solution. This solution must be temporally consistent,
semantically consistent, fulfilling the constraints of the formal model, and explaining
the findings of the temporal database information.
Practical Experience.
This paper also relates our practical experience in the application of this architecture.
To this end, we choose the medical domain, in particular the Intensive Care Unit (ICU)
one.
Figure 1. ACUDES: Architecture for intensive Care Unit Decision Support.

Physicians at intensive care units have to deal with an overwhelming amount of data
provided not only by on-line monitoring but also collected from patients’ records (e.g.,
laboratory results), which are, in most cases, collected manually at different time in-
stants. In order to provide efficient decision support systems and medical research tools
in the ICU domain, it is necessary to integrate and analyse the information provided
from these different sources. The result is ACUDES , the Architecture for intensive
Care Unit Decision Support.
ACUDES is a specific architecture that follows the proposed structure of this pa-
per (figure 1). Thus, it includes the Temporal/Domain Module. This module contains
the FuzzyTIME temporal reasoner and the Ontology Server, where the ICU ontology is
allocated. Furthermore, ACUDES also includes the Decision Support Module. There,
it is worth noting that we have designed and implemented CATEKAT, a web-based
tool to acquire the domain knowledge for the diagnosis model. Finally, the core of the
ACUDES diagnosis process is described by an abnormal behaviour model-based algo-
rithm. This model describes the causal and temporal behaviour of diseases. Thus, the
algorithm, from an abductive approach, obtains a diagnosis solution given the symp-
toms and the temporal behaviour model.
References
1. Console, L., Torraso, P.: A spectrum of logical definitions of model-based diagnosis. In
Hamscher, W., Console, L., de Kleer, J., eds.: Readings in Model-Based Diagnosis. Morgan
Kauffmann Publisher, Inc. (1992) 78–88
2. Campos, M., A.Cárceles, Palma, J., R.Marı́n: A general purporse fuzzy temporal information
management engine. In: Proceedings of the EurAsia-ICT 2002. (2002) 93–97
Ontology–Based Models of Human–Computer
Interaction for Usability Evaluation and Usable
Design
Elena Garcı́a-Barriocanal and Miguel-Ángel Sicilia
Computer Science Department. Polytechnic School.

University of Alcalá. Ctra. Barcelona km. 33.6. 28871 – Alcalá de Henares (Spain)
elena.garciab@uah.es,msicilia@uah.es
Abstract. The field of Human–Computer–Interaction (HCI) is concerned

with the design and evaluation of user interfaces, and takes a pluri-
disciplinary stance that has resulted in a diversity of models, techniques
and tools. Formal ontologies can be used as a mean for achieving a shared
understanding and representation of such artifacts, thus enabling com-
mon tools that enhance key processes in both evaluation and design.
This paper describes a description logics ontology of HCI and several
associated intelligent information processing functions oriented to assist
developers in HCI–related tasks. Concretely, the use of a representation
of design guidelines to implement design advice and the semi–automated
design of evaluation tools are described as applications. The ontology de-
scribed is intended as a general framework for Semantic Web applications
related to considering user interaction and human factors.
Keywords. Ontology, Human-Computer Interaction, User Interface Evalu-

ation, User Interface Design, Design Guidelines.
1 Introduction and Motivation

The field of Human–Computer–Interaction (HCI) is concerned with the design
and evaluation of user interfaces, and takes a pluri-disciplinary stance that has re-
sulted in a diversity of models, techniques and tools clearly reflected in proposed
curricular models1 . This makes specially difficult the development of compute–
aided tools for user interface design and calls for rich and flexible knowledge
representation structures that codify practices and knowledge of tools, models
and techniques. Formal ontology [1] has been used recently for the modelling of
some design aspects [5], but more areas have to be covered to provide a frame-
work for both analysis and design. This paper describes a first effort towards such
ontological structure, and provides the description of case studies regarding the
application of information processing based on the ontology to the design and
evaluation aspects. The ontology described is intended as a general framework
1
http://www.acm.org/sigchi/cdg/cdg2.html
for Semantic Web applications [2] related to considering user interaction and
human factors.
2 Basic Ontological Structure
The work described here is founded on the analysis described by van Welie et al
[8], which establishes three knowledge domains: task domain, design domain and
user domain. In the basis of these domains and considering the usability defini-
tion proposed by ISO [3], we have structured the current version of our ontology
in six integrated (sub-)ontologies (see Figure 1). Aspects and usage indicators
of usability are described in the ’upper’ core concepts ontologies, and concepts
related with them are structured in two general partitions, one regarding the es-
sential interest point of usability, i.e. human-system interactions, and the other
regarding activities, methods and tools that are (left and right parts of Figure 1,
respectively). In consequence, concepts describing the physical fact of the inter-
action are separated from normative constructs about evaluating or designing-for
usability. The former include the application domain describing the structure of
the software-hardware application or system, the constructs describing human
users, their profiles and perceptual traits, and their purposeful interactions or
tasks, consisting ultimately on actions carried out by the users on the ’affor-
dances’ provided by the system. The latter includes norms and advices for the
design of user interfaces, and also methods, techniques and interpretations of
the way to evaluate and improve those designs, as a part of a usability-oriented
software development process or cycle. The application domain describes the
Fig. 1. Overall view of the HCI ontology
conceptual and physical structure of the user interface, and the human domain
describes the users and their abilities and cognitive processes. The task domain
models the dialogue and interaction between humans and systems, at different
levels of detail. The application domain is of special interest in some cases, e.g.
when considering mobile devices, due to the heterogeneity of their features that
require specific studies. The FIPA ontology2 provides a foundation for that kind
of specific domain. Design and evaluation are two interacting activities that serve
the purpose of making more ’usable’ interactions.
3 Usability Evaluation using Ontological Support

Following the case study and prototype system described by Garcı́a et al. [5],
an ontology of evaluation has been structured around the following core con-
cepts: (a) application mereology, (b) evaluation techniques and artifacts, and (c)
guidelines or specialized knowledge required.
Application mereology refers to the structure of parts of the application under
consideration, in some cases reflecting also contextual or organizational issues.
Such structure can be described at various levels, up to the fine granularity of
markup elements and attributes in the case of Web systems. An example of a
coarser grain is the following:
W AT ype v ∀usuallyIncludes.W AP artT ype
W AP artT ype(RecommendationSystem); W AP artT ype(RegP age)
W AP artT ype(SearchP age); W AP artT ype(CartP age)
usuallyIncludes(e − shop, RecommendationSystem)
usuallyIncludes(e − shop, CartP age)
usuallyIncludes(e − shop, RegP age)
usuallyIncludes(e − shop, SearchP age)
In the example, applications are described in terms of ‘typical parts’, which

are specially useful for heuristic inspection or questionnaire evaluation, in which
some expert knowledge is codified in terms of guidelines or general recommen-
dations.
Evaluation techniques and artifacts are represented inside the ontology, in-
cluding its process of application, which supports the construction of semi-
automated guiding tools for usability engineers. Finally, guidelines and special-
ized knowledge are encoded in the form of classical if-then rules (with the in-
clusion of vagueness in some cases [4]), and also as general ontological frames
representing complex patterns as those prescribed in information architecture
texts.
4 Assisting Design with Semantic Web Technologies

Assisting design with advice and contextual evaluation has been addressed by
existing tools, e.g. [6]. The ontological structure described enhances existing tools
in two directions. First, it allows a generalization of existing representations with
more flexible and open languages. Second, ontologies combined with Web Service
architecture (e.g. as described in WSMO3 ) enables the construction of usability-
related knowledge bases available through the Web. This paper describes such
a representation and structure, extending existing schemas [7].
2
FIPA Device Ontology Specification, doc. number XC00091C, available at
http://www.fipa.org.
3
http://www.wsmo.org/
The main extensions include the representation of explicit degrees of evi-
dence (modelled after the notions in an existing service 4 ), and the provision of
a hierarchy of degrees of requirement, including obligatory, recommended and
suggested aspects. This allows the specification of compliancy requirements like
those of accessibility:
M arkupElement v ∀requires.M arkupAttribute
M arkupElement(ImgHT M LElement); M arkupAttribute(AltAttribute);
requires(ImgHT M LElement, AltAttribute)
But it also allows for the specification of overall recommendations on the struc-
ture of applications through inference, e.g.:
LongHierarchyStructure ≡ HypermediaStructure u
∃linksT o.(HypermediaN ode u ∃linksT o.(HypermediaN ode u ∃linksT o.HypermediaN ode))
This simple example on hierarchy depth illustrates the capabilities of description–

logics classification to reason on application structure while design proceeds.
Such axioms combined with operational rules that trigger (possibly remote)
classification as a result of some user actions with the design tools result in
a design–advice layer that can be updated by simply extending the ontology.
In addition, the above described rich mereology and explicit representation of
context enables diverse levels of application of design knowledge.
References
1. Baader, F., Calvanese, D., McGuinness, D., Nardi, D., Patel-Schneider, P. (eds.).
(2003). The De-scription Logic Handbook. Theory, Implementation and Applica-
tions, Cambridge.
2. Berners-Lee, T., Hendler, J., Lassila, O. (2001). The Semantic Web. Scientific Amer-
ican, 284(5), pp. 34-43.
3. Bevan, N. (1998) ISO 9241: Ergonomic requirements for office work with visual
display terminals (VDTs) - Part 11: Guidance on usability. TC 159//SC 4 Technical
Committee of International Organization for Standardization.
4. Garcı́a, E., Sicilia, M. A., Gutiérrez, J.A. On the Vague Modelling of Web Page Char-
acteristics Regarding Usability. Atlantic Web Intelligence Conference (AWIC03),
Madrid, Spain. LNCS 2663. Springer, pp. 199-207. (2003).
5. Garcı́a, E., Sicilia, M.A., Hilera, J.R., González, L. Dialogue-Based Design Of Web
Usability Questionnaires Using Ontologies. Computer Aidded Design of User Inter-
faces IV, Kluwer Academic Press (to appear).
6. Scapin et al. (2000). A Framework for Organizing Web Usability Guidelines. 6th
Conference of Human Factors on the Web, Austin, TX.
7. Vanderdonckt, J., Development Milestones towards a Tool for Working with Guide-
lines, Interacting with Computers, Vol. 12, N2, 1999, pp. 81-118.
8. Van Welie, M., van der Veer, G.C. & Eliëns, A. (1999), Breaking down usability,
Proc. of Interact’99. 613-620.
4
http://usability.gov/guidelines/
Extension of Ontologies assisted by Automated
Reasoning Systems?
Joaquı́n Borrego-Dı́az?? , Antonia M. Chávez-González

Computational Logic Group
Departamento de Ciencias de la Computación e Inteligencia Artificial.
E.T.S. Ingenierı́a Informática-Universidad de Sevilla.
Avda. Reina Mercedes s.n. 41012-Sevilla
Extended Abstract
The Ontology management has becomed in a critical issue in fields re-

lated with Knowledge Representation (KR) and Intelligent Information
Processing as the Semantic Web. An important feature of the problem is
the need of extending or revising ontologies. In general, such task is, from
the companies point of view, dangerous and expensive: every change in the
ontology affects to the overall knowledge organisation of them. Moreover,
it is also known that the self process of extension is hard to automatize:
the tools are designed to facilitate the syntactic extension or ontology
mapping. But the effect of Ontology mapping on the logical reasoning is,
in general, unknown. Particularly hard is to analyse the semantic conse-
quences of ontology mapping on the use of automated reasoning systems
[3].
The aim of this paper is to propose a practical definition of robust-
ness for ontologies (following the foundational principles given in [3] and
suggested by the computer assited cleaning of Knowledge Databases [1])
as well as a ARS-assisted method to extend ontologies which produces
robust ontologies (according to our definition). The definition is more
manageable than logical notions as categoricity or completeness.
With respect to the method, it basically consists in four steps aided
by the automated theorem prover OTTER and its complementary au-
tomated model finder MACE4. In order to obtain a practical method,
some of the basic (theoretical) logical principles required by the defini-
tional methodologies of building of formal ontologies must be weakened
[5]. Starting with a robust ontology,
?
Partially supported by TIC-137 of Plan Andaluz de Investigación and by the project
TIN2004-03884 of Spanish Ministry of Education and Science, cofinanced by FEDER
founds.
??
Corresponding author. jborrego@us.es
1. First, one extends the lattice of the basics concepts of the ontology
by certain equational characterization of the lattice.
2. Next, one applies MACE4 on a possible axiomatization of the new
lattice in order to obtain the new lattices. In general, the characteri-
zation of the lattice is a theory weaker than the initial ontology.
3. The third step consists of the refinement of the axiomatization (by
adding of new equations) in order to MACE4 exhibits one only one
model (that is, the axiomatization is categorical).
4. Finally, it is necessary to certify (by means OTTER or hand-made)
the unicity of above model.
The final stage of the method is not algorithmic. It consists in an on-
tological interpretation of the new relation, re-interpreting (generally by
refining) if necessary, the older ones. This task, nonalgorithmic in nature,
is responsability of experts on the universe that is represented by the
ontology.
The abstract presentation of the extension method is applied in an
interesting case, the extension of ontologies on Qualitative Spatial Rea-
soning under insertion of relations on imperfect spatial information. Con-
cretely, we analyse the extension of the well-known Region Connection
Calculus [4], a theory which can be used as a meta-ontological tool for
analysing ontological anomalies [6] as well as cleaning them [2].
References
1. J. A. Alonso-Jiménez, J. Borrego-Dı́az, A. M. Chávez-González and J. D. Navarro
Marı́n, Towards a Practical Argumentative Reasoning with Qualitative Spatial
Databases, 6th International Conference on Industrial and Engineering Applica-
tions of Artificial Intelligence and Expert Systems, Lecture Notes in Artificial In-
telligence vol. 2718, 789-798, Springer-Verlag Berlin, 2003.
2. J. A. Alonso-Jiménez, J. Borrego-Dı́az and A. M. Chávez-González, Ontology
Cleaning by Mereotopological Reasoning. Proceedings of DEXA Workshop on Web
Semantics (WebS’04). IEEE Press, pp. 137-137, 2004.
3. J. A. Alonso-Jiménez, J. Borrego-Dı́az and A. M. Chávez-González, Foundational
challenges in Automated and Ontology Cleaning in the Semantic Web. To appear
in IEEE Intelligent Systems.
4. A. G. Cohn, B. Bennett, J. M. Gooday and N. M. Gotts. Representing and Rea-
soning with Qualitative Spatial Relations about Regions. Chapter 4 in O. Stock
(ed.), Spatial and Temporal Reasoning, Kluwer, Dordrecth, 1997.
5. B. Bennett, The Role of Definitions in Construction and Analysis of Formal On-
tologies, in: P. Doherty, J. McCarthy, and M. Williams (eds.) Logical Formalization
of Commonsense Reasoning (2003 AAAI Spring Symposium), 27-35, AAAI Press,
2003.
6. A.M. Chávez-González, Mereotopological Automated Reasoning for Ontology
Cleaning, forthcoming Ph.D. thesis.
Applying Genetic Algorithms to Referring Expression
Generation and Aggregation
Raquel Hervás and Pablo Gervás
Departamento de Sistemas Informáticos y Programación,

Universidad Complutense de Madrid, Spain
raquelhb@fdi.ucm.es,pgervas@sip.ucm.es
1 Introduction
Natural Language Generation (NLG) includes complex tasks [2], and features such as
ambiguity can make these tasks too complex in fact to be tackled with classic search
algorithms in a reasonable time. Evolutionary Algorithms (EAs) are heuristic tech-
niques [1] based on a search model which emulates evolutionary phenomena, and that
can be used to speed up the search. The purpose of most EAs is to find a good solu-
tion but not necessarily the best solution, and this is enough for most natural lan-
guages processes. In this paper we present a first approach to the idea of using NLG
and EAs together.
The specific problems addressed in this paper are two subtasks of the process of
automatically generating a natural language text for a given content: referringf expre-
sion generation and aggregation. The nature of these subtasks is best illustrated over
an example:
A princess lived in a castle. She loved a brave knight. She was
pretty. The castle had towers. It was strong. They lived in it.
In this text various elements of the reality that is being described are mentioned
one or more times: a princess, a knight,... Different natural language expressions are
used to describe them at each occurrence in the text: “a princess”, “she”, “the prin-
cess”. The task of referring expression generation concers the way in which a system
decides at each occurrence how to refer to a given element.
The correct use of referring expressions to compete with human generated texts
involves a certain difficulty [4]. Possible simple algorithms for deciding when to use
a pronoun and when to use the full name produce poor results. Two occurrences of
the same concept in a paragraph can be far apart, and this may confuse the reader.
Knowledge intensive approaches modelled on the way humans do it require a certain
measure of content understanding that is resource hungry.
Aggregation [3] involves deciding how compact the presentation of information
should be in a given text. It operates at several linguistic levels, but we only consider
it here with respect to concepts and their attributes. For instance, the system must
decide between generating “The princess is blonde. She sleeps.” and generating “The
blonde princess sleeps.”. Aggregation is generally desirable, but may result in adjec-
tive-heavy texts when the information to impart becomes dense in terms of attributes,
as in “The pretty blonde princess lived in a strong fancy castle with her stern rich
parents.”. A good solution should find a balance between conciseness and acceptable
style.
2 An Evolutionary Solution
We propose the use of a simple genetic algorithm, where the appearances of the con-
cepts are considered as the genes. The initial population is generated randomly, using
for each concept its full name or its pronoun. When using the full name, a selection of
the attributes the concept has in the knowledge base is chosen. These attributes will
appear just before the name of the concept, as it is usual in English. The system
works over this population for a number of generations determined by the user. In
each generation three genetic operators are used: crossover, mutation and aggrega-
tion. Finally, at the end of each generation each tale is evaluated and a selection of the
population is passed to the next one, in such way that the tales with a higher fitness
have more possibilities of being chosen.
The operators
For the crossover operator, two drafts are selected randomly and crossed by a ran-
dom point of their structure. So, each of the sons will have part of each of the parents.
In the case of the mutation operator, some of the genes are chosen randomly to be
mutated. If the gene is a pronoun, it will change into the correspondent full name,
always associated with a subset of its possible attributes. In case the gene was a full
name, there are two options: to change it into a pronoun, or to change the subset of
attributes that appear with it. One of these two options is chosen randomly.
The aggregation operator addresses the task of deciding on the aggregation be-
tween concepts and their attributes. This involves a certain modification of the struc-
ture of the text, because sentences in the text may be deleted if the information they
impart becomes part of a previous sentence. The aggregation operator acts only on
genes corresponding to explicitly mentioned concepts: concepts referred by pronouns
are excluded. It can act in two directions. If the reference to the concept appears with
one or more attributes – as in “A blonde princess lived in a castle.” -, the operator
disggregates the attributes by eliminating their mention and adding a corresponding
“X is Y” sentence – resulting in “A princess lived in a castle. She was blonde.”. If the
reference to X has no attributes – as in “A princess lived in a castle.” -, the algorithm
looks for an “X is Y” sentence – such as “The princess was blonde.”-, adds the cor-
responding attributes to the reference, and deletes the “X is Y” sentence – resulting in
“A blonde princess lived in a castle.”.
The goal of this definition of the aggregation is to ensure that the attributes of a
concept be mentioned in the appearance of a concept or in the correspondent “X is Y”
sentences, but not in both. As the aggregation operator are used randomly, the desired
result is obtained only in some cases.
The fitness function
The key to the genetic algorithm lies in the choice of fitness function. A simple ap-
proach would be to require that in each generation the user reads all the texts and
gives them a fitness value. The number of generations and individuals in the popula-
tion for a simple experiment makes this approach impractical. So, we analyzed the
features of a human generated text from the point of view of the referring expres-
sions, and we found five different evaluation functions, each one aimed at a different
aspect.
Correct Referent. When writing a text, we cannot use a pronoun for something
that we have not mentioned before, or readers would get confused. In addition, if the
full name reference and the pronoun are far, the reader can also get confused and be
unable to link the two occurrences of the same concept. So, this function penalizes
the appearances of pronouns for a concept that has not been referred in the two previ-
ous genes with its full name.
Redundant Attributes. When describing a concept in an “X is Y” sentence, peo-
ple do not use the attribute they are going to describe in the reference to the concept.
This function penalizes sentences like “The blonde princess was blonde”.
Reference Repetition. Using always the same reference together with the same at-
tributes results in repetitive text. For example, it is acceptable to use “the princess”
every time we refer to the princess character, but it would be striking to use always
“the pretty blonde princess”. To avoid that, repetitive use of references is penalized.
Coherence. If we use different subsets of its attributes in different references to
the same concept, the reader may mistakenly assume that we are referring to different
concepts. For example, if we use “the pretty princess” and “the blonde princess” in
different places, and we have not specified that the princess is both pretty and blonde,
it could seem that there are two princess: a pretty one and a blonde one. With this
function, we penalized the appearance of attributes that have not been mentioned in
the first description of a concept or in a “X is Y” sentence.
Lack of Concept Information. When applying the genetic operators to a draft,
some information about a concept may disappear from the final text. To avoid that,
this function looks for all the attributes of each concept in the ontology, and checks if
they are mentioned in the text. If not, the draft is penalized.
For the evaluation of each of the texts that form the population, we use formula
(1), where error is the accumulated error of a specific text taking into account the
fitness functions mentioned above.
fitness = 1 / ( error + k ). (1)
In this way, the fitness would be greater if the error is smaller. The constant k is
used to avoid divisions by zero. In our experiments it was set with the value 1, so the
maximum possible fitness was 1.
The results
The following text is an example of system output for a very simple input concerning
the initial description of the characters that take part in a tale:
A princess lived in a castle. She was the daugther of parents. They
were stern. The blonde princess loved a brave handsome knight. The cas-
tle had towers. The parents lived in the castle. It was strong.
The beginnings of stories are rich in descriptions of characters and locations and
are therefore good case studies for the problems being discussed. The input to the
system provides conceptual descriptions of characters and locations in terms of at-
tributes and relations between them. The evolutionary algorithm operates over the
output of previous stages that have carried out other NLG tasks such as content de-
termination – selecting the particular concepts that are relevant – and discourse plan-
ning - organising them in an orderly fashion. These tasks are currently carried out in a
traditional manner and simply provide the data for the evolutionary stages. A tradi-
tional surface realization stage transforms the output of the evolutionary stages into
readable text.
3 Conclusions and future work
With respect to both of the tasks addressed, the output texts respect the strong con-
straints required for the text to be acceptable, while at the same time showing reason-
able variation between the different options much as a human-generated text would.
Due to the main difficulties in the evaluation of the population at the end of each
generation, we are currently considering the possibility of using neural networks to
reproduce the human evaluation of the quality of the resulting texts. We are also
working on extending the system to allow the use of proper nouns to describe some
concepts, as an additional option to pronouns and descriptive references, including
the revision of the genetic operators and the introduction of new evaluation functions
to estimate the correct application of proper nouns. In view of these results, in future
work we want to apply EA techniques to other tasks of NLG, such as content deter-
mination and discourse planning. The particular advantages of evolutionary tech-
niques, combined stage by stage in this manner, may be an extremely powerful
method for solving natural language generation problems while also profiting from
classic NLG techniques.
References
1. Holland, J. H., Adaptation in Natural and Artificial Systems. An Introductory Analysis with
Applications to Biology, Control and Artificial Intelligence, MIT Press, Cambridge, Massa-
chusetts, Second Edition, 1992.
2. Reiter, E. “Has a consensus NL generation architecture appeared, and is it psychologically
plausible?”, in: McDonald, D. and Meteer, M. (eds.) Proccedings of INLGW ’94, pages
163-170, Kennebunkport, Maine, 1994.
3. Reape, M, Mellish, C., “Just what is aggregation anyway?”, in: Proceedings of the 7th Euro-
pean Workshop on Natural Language Generation, Toulouse, France, 1999.
4. Reiter, E. Dale, R. “A fast algorithm for the generation of referring expressions”, in: Pro-
ceedings of the 14th conference on Computational linguistics - Volume 1, Nantes, France,
1992

!"# $ % % %
& ' (# & ) %
* +# ' ) '
$ ) %
# %
Æ # , -#
, - & %

# . ' $
# , /
' & ) &
#
& 0" 1"# )+
Æ ' # 2
* $ ) %
' # , & $ %
) $ # %
& ' $ Æ #
)+ $ % $
' & % ) # ,
' $# , )+
3" 4"&
#
$ ) %
' $ )

#
, 5 $ 6

$ $ $
#
& %
' & $ # , ,2
$ $
) $
$ ) # , $7 '

( % $ '
) ) #
8 + Æ & $
# 9 & )+
:" * $ '

# *
Æ $
# ;) & $
$ #
) ) +
$ ) & ) /
7 $ ) # - &
$ (
$ ) ( # ,
& ) # ,
$ & 4"& $
) $ #
)+& ) ) %
$ # < ) $
& #
%
* ' *
, .= & ) +#

# , )+ ) % $ >? %
= , @> -.. 7 ,=!AA!%
AAB4C D E 7 2E,A12D=3A:A32F#

!!! "#$%
"& $''$
$ (
)*+ (,!%" ,
!
" #$
# %
- ..#
" /0 (
& 1
2 $''$
# 3%4 (%5 6 6%6 %7
8 %

8
'
".)$''"*+9%" 6
$''"
& 5 : , 7 8
%

( &
&' ) #
# $

#$ "'&%"$ $''#
; < 3 /0 ( ' )*
" $ / 1
7

5 ...
An Ontology for Reusing Synthetic Tasks
Abraham Rodríguez-Rodríguez, Francisca Quintana-Domínguez
Dept. of Computer Science

University of Las Palmas de Gran Canaria
Edificio de Informática. Campus de Tafira
35017 Las Palmas de Gran Canaria
Canary Islands- Spain
{arodriguez, fquintana}@dis.ulpgc.es
Extended Abstract
Knowledge-Based Systems (KBS) development must overcome multiple difficulties,

making it a high cost technology with a large number of development failures. Cur-
rent development methodologies, such as VITAL [4], MIKE [1] or CommonKADS
[8], stress the use of generic components to facilitate the reuse of knowledge, Prob-
lem Solving Methods (PSM) and code.
Ontologies play an important role in this process because they try to normalize the
task and/or the domain they that describe [3]. However, as discussed in [5], there are
some gaps between the PSM and the task it carries out, on one side, and between the
PSM and the knowledge it applies, which complicate domain reuse.
This paper will describe our experience on the development of some synthetic ap-
plications (see Table 1), how the same PSM was applied to four domains, and how
the different domain knowledge corpus were mapped onto the PSM ontology.
Table 1. We use the task types as they are defined in the CommonKADS’s book [8]
Application Synthetic Task Type

Educational centre timetable Scheduling
Employee timetable Scheduling
PC configuring system Configuration design
Airline lines programming Planning
CommonKADS and authors like [6] suggest a shared characterisation of all syn-
thetic tasks by using some terms such as requirements, constraints, domain, etc. These
terms are similar to those used by the Constraint Satisfaction (CS) techniques [2],
such as the Forward Checking algorithm. We have adopted this technique (Figure 1
shows de derived PSM), because it provides us with the needed simplicity to trans-
form it into a PSM and is easily extensible with some other techniques such as consis-
tency techniques and variable/value ordering. A detailed description of this PSM can
be found in [7].
The full paper will describe the roles of the PSM (i.e. task description knowledge,
soft requirement, etc.) and the association between the domain vocabulary and some
of the test applications.
According to the people participating in this initiative, the reuse of the PSM and its
associated executable code supposed at least 30% of development effort saving (just
for the expert module). This percentage is even larger depending on some other fac-
tors such as the complexity of the application being modelled. The PSM Ontology
allow us to prepare structured interviews with the human experts in order to acquire
specific knowledge for that role. The experts easily understood the PSM diagram and
collaborated actively providing valid rules that were later included in the system.
Task
Input Data Generate Composition
Knowledge
Task
Opera- Partial
Description
tionalize Solution
Knowledge
Soft
Requirement
Propose Undo
Extension Extension
Preferences
Ordering
Knowledge
Extension Violation
Hard Check Valid

Requirement Forward Solutions
Fig. 1. Forward Checking PSM
On the other hand, assessing a priori the potential benefits of reusing components
is not easy because there are too many (obscure) factors that directly influence on the
final revenue. Some of these include the technological background and experience of
the development team, the size and complexity of the task type and domain applica-
tion, and how easy the adaptation of the module will result to be reused to the actual
domain problem.
References
1 Angele, J., Fensel, D., Studer, R. Domain and Task Modelling in MIKE. In Suthliffe, A et al.
(eds). Domain Knowledge for interactive system design. Chapman & Hall. 1996
2 Baptiste, P., Le Pape, C., and Nuijten, W., Constraint-Based Scheduling. Applying con-
straint programming to scheduling problems. Kluwer Academic Publishers. 2001
3 Chandrasekaran, B., Josephson, J., Benjamins R. Ontologies: What are they? Why do we
need them? IEEE Intelligent Systems and Their Applications, 14(1): 20-26, 1999. Special
issue on Ontologies.
4 Domingue, J., Motta, E., Watt, S. The emerging VITAL workbench. In Aussenac et al.
editor, EKAW’93 Knowledge Acquisition for Knowledge-Based Systems. Lecture Notes in
Artificial Intelligence, LNCE723, Berlin Germany, 1993, Springer-Verlag.
5 Gómez-Pérez, A., Benjamins, R. Overview of knowledge sharing and reuse components:
Ontologies and Problem-Solving Methods. Proceedings of the IJCAI-99 workshop on
Ontologies and Problem-Solving Methods (KRR5) Stockholm, Sweden, 1999
6 Motta, E., Rajpathak, D., Zdrahal Z., Roy, R. The Epistemology of Scheduling Problems.
Proceedings of the 15th European Conference on Artificial Intelligence, F. van Harmelen
(editor), IOS Press, Amsterdam, 2002
7 Rodríguez, A., Palma, J., Quintana, F. Experiences in Reusing Problem Solving Methods.
An application to Constraint Programming. Knowledge-based intelligent information and
engineering systems. LNAI 2774, Springer, 2003.
8 Schreiber, Akkermans, Anjewierden, de Hoog, Shadbolt, Van de Velde, Wielinga. Knowl-
edge Engineering and Management. The CommonKADS Methodology. The MIT Press.
1999
Modelling and Simulation of Systems with
Multi-level Anticipation and Retardation
Daniel M. Dubois1 and Stig C. Holmberg2

1
U.E.R. Operations, HEC, rue Louvrex, 14, B-4000 Liège, Belgium
CHAOS asbl, Institute of Mathematics, University of Liège, Grand Traverse 12,
B-4000 Liège, Belgium
Daniel.Dubois@ulg.ac.be,
WWW home page: http://www.ulg.ac.be/mathgen/CHAOS
2
Dept. of Information Technology and Media, Mid Sweden University
S-83125 ÖSTERSUND, Sweden
shbg@ieee.org,
WWW home page: http://www.itm.miun.se/~stihol/
1 A Management Problem and its Knowledge Challenges

In practice managers often anticipate future situations for taking decisions at
the present time. Mathematical models proposed in the literature, however, very
often do not take this anticipatory dimension into account. Contrary, the models
at hand merely represent the dynamics at the present current time. At best, some
of those models extend these current-time dynamics to include retardation.
The purpose of this paper is hence to demonstrate that it is possible to
extend the classical modelling techniques to include anticipative capabilities. We
will base our work on a general management problem dealing with multi-level
actions and decision-making on operative, strategical, and normative planning
levels (Espejo et al, 1996).
Espejo et al (1996)have with their multi-level model highlighted the decision
dilemmas of managers. Holmberg (1999) has already identified the anticipating
power of the model. This means that an action on one level will not only have an
effect on that level in focus. With some retardation (delay) it will also have effects
on the others. This, of course, is quite in line with classical system knowledge.
Those indirect and often dramatic effects are anyhow often overlooked in many
practical situations with the manager in charge just focusing on the direct effects.
The approach of Espejo et al (1996) is a good start for ameliorating current
decision support for managers but as the model is expressed only in qualita-
tive and graphical form it is not directly suitable for computer simulations and
quantitative calculations. However, by strengthening this basic model with an-
ticipatory computing techniques (Dubois, 2003) we will be able to advance the
state of the art.
2 The Proposal and its Assessment

In showing the effects of retardation and anticipation in business models, let us
consider the following recursive discrete model.
P (t + 1) = P (t) + dtG1 [P (t), F (t), V (t); p1 ] (1)
F (t + 1) = F (t) + dtG2 [P (t), F (t), V (t); p2 ] (2)
V (t + 1) = V (t) + dtG3 [P (t), F (t), V (t); p3 ] (3)
where dt is the temporal time step.
The state of the first operational level is given by the temporal function P(t),
which may represent the production of the company. The state of the second
strategic level is given by F(t) representing, for example, Research and Devel-
opment. The state of the third level, at last, is given by V(t), representing the
normative work of defining the company’s vision and mission. The three func-
tions Gi are functions of the three states P (t) , F (t) and V (t) and the sets of
parameters given by pi .
The model as it is, is a model describing the company at the current time
t. The numerical simulation of this model can be performed by giving explicitly
the functions Gi , the initial states at time t = 0 for P (0) , F (0) and V (0) and
the values of the parameters pi . Let us now give an example of retardation and
one of anticipation:
P (t + 1) = P (t) + dtG1 [P (t), F (t − τ1 ), V (t); p1 ] (4)
F (t + 1) = F (t) + dtG2 [P (t), F (t), V (t); p2 ] (5)

V (t + 1) = V (t) + dtG3 [P (t + τ2 ), F (t), V (t); p3 ] (6)
where τ1 is a factor of retardation and τ2 is one of anticipation.
The simulations show different behaviour, varying from stable fixed point to
completely chaotic, depending on the values of the parameters. The introduction
of retardation leads to a less stable solution while the effect of the anticipation
is to render the system more stable. This approach can also be applied in other
fields of research and/or development in computer supported systems theory.
References
1. Dubois Daniel M.:Mathematical foundations of discrete and functional systems with
strong and weak anticipations. In Anticipatory Behavior in Adaptive Learning Sys-
tems, State-of-the-Art Survey, Edited by Martin Butz et al, Lecture Notes in Ar-
tificial Intelligence, Springer LNAI 2684. (2003) 110–132
2. Espejo Raul, Schuhmann Werner, Schwaninger Markus, Bilello Ubaldo: Organi-
zational Transformation and Learning, A Cybernetic Approach to Management.
Wiley, Chichester. (1996) 227–270
3. Holmberg Stig C.:Multimodal Systemic Anticipation. I. J. of Computing Anticipa-
tory Systems, Vol. 3. (1999) 75–84
A Tractable Subclass of Fuzzy Constraint Networks
Alfonso Bosch1, Francisco Guil1, and Roque Marin2
1 Dpto. de Lenguajes y Computacion

Universidad de Almeria
04120 Almeria (Spain)
{abosch, fguil}@ual.es
2 Dpto. de Ingenieria de la Informacion y las Comunicaciones

Universidad de Murcia
Campus de Espinardo
30071 Espinardo (Murcia, Spain)
roque@dif.um.es
Abstract. The Fuzzy Constraint Networks model, a generalization of the

Disjunctive Temporal Fuzzy Constraint Networks, is a framework that allows
representing and reasoning with fuzzy qualitative and quantitative complex
constraints. However, its general complexity is exponential, and we need to find
tractable subclasses. In this paper we propose two algorithms to deal with a
tractable subclass named Series-Parallel Fuzzy Constraint Networks.
1 Introduction
Fuzzy Temporal Constraint Networks model (FTCN), introduced in [4], allows

expressing simple constraints, representing them with a possibility distribution. Fuzzy
temporal constraints allow combining precise, imprecise, qualitative and quantitative
information. A fuzzy model allows intermediate consistency degrees, and to quantify
the possibility and necessity of a relationship or query.
In certain tasks, such as planning or scheduling, a more general model with
disjunctions is needed. Then, the FTCN model is enhanced, defining a constraint with
a finite set of possibility distributions, obtaining the Disjunctive Fuzzy Temporal
Constraint Networks (DFTCN) [1]. This model extends the TSCP framework
proposed by Detcher [3], and it allows constraints as “Irrigation is much before or a
little after than Treatment”, and subsumes the Vilain & Kautz Point Algebra. This
framework allows representing all the possible relationships between time points,
between intervals and between time points and intervals, and their disjunctions.
The main drawback of DFTCN is its computational inefficiency. These networks
are non-decomposable, needing backtracking to find a solution [1]. Determining the
consistency and computing the minimal network are also exponential, and it would be
interesting to find tractable subclasses, as series-parallel networks [2].
In this paper we use a generalization of DFTCN, named Fuzzy Constraint
Networks (FCN) [2], and we present two efficient algorithms, SP+ and SPR, to decide
if a FCN is consistent, series-parallel and obtain its equivalent path-consistent (and
minimal) network, using the information of the original problem constraints.
2. Series-Parallel Fuzzy Constraint Networks
A Fuzzy Constraint Network (FCN) Ld consists of a set of n+1 variables X0, ... ,Xn
whose domain is a full ordered set of equidistant precise values, and a finite set of
disjunctive binary constraints among these variables. The variables and constraints
can be represented with a constraint graph or network. The complete model definition
is included on [2].
A network is series-parallel if it can be reduced to an arc applying iteratively a
reduction operation. The SP algorithm [2] decides whether a network is series-
parallel, using only the graph topology, but not the information contained into
network constraints. Enhancing this algorithm to use constraint information, we can
obtain useful information about the network, and to decide simultaneously whether
the network is series-parallel or not.
Working in this direction, we propose the algorithm SP+, which can decide with an
O(n) complexity whether a network is inconsistent if it detects an empty constraint. If
an empty constraint is not detected and the algorithm decides that the network is
series-parallel, the network is consistent. As a subproduct, the final obtained
constraint is the minimal constraint among the two lasting variables, and the
equivalent network obtained with the computed constraints is decomposable in the
reverse variable reduction ordering. This is the code of SP+ algorithm:
SP+ Algorithm
begin
for each i=0..n Compute-degree (i)
NodeQueue = {nodes with degree 1 and 2}
while (NodeQueue <> ∅ and |V| > 2)
begin
node=Get(NodeQueue)
if Degree(node) = 2 then Lij ← Lij ∩
&Lik ⊕
&L
kj
if Lij = ∅ then exit
V ← V − {node}
if Degree(node)=1 then Degree(Neighbour(node)) --
if Degree(Neighbour(node)) = 2 then
Put(NodeQueue,Neighbour(node))
else if Connected(Neighbours(node)) then
Degree(Neighbour(node)) --
if Degree(Neighbours(node)) = 2 then
Put(NodeQueue, Neighbours(node))
else E ← E + {ArcNeighbours(node)}
end
if (NodeQueue = ∅ and |V| > 2) then
exit (“Network is not series-parallel”)
end
Fig. 1. Example of network to be reduced with SP+
Figure 1 shows a FCN network to be reduced with SP+. Figure 2 shows the
reduction process and refined constraints obtained.
Fig. 2. Example of reduction process

Based on the results of SP+, we can propose a linear rebuilding algorithm SPR,
which obtains a path-consistent network respecting the original network topology.
The obtained network is not the full path-consistent or minimal network, because it is
not a complete graph.
As future work, we can highlight the evaluation of these techniques and its
comparation versus other approaches for series-parallel networks, the implementation
of a linear algorithm for obtaining the minimal network, and the search of another
FCN tractable subclasses and decomposition methods.
References
1. A. Bosch, M. Torres, and R. Marín. Reasoning with Disjunctive Fuzzy Temporal Constraint
Networks. In Proc. of 9th Int. Symposium on Temporal Representation and Reasoning
(TIME’02). IEEE Computer Society Press, pages 36-43, 2002.
2. A. Bosch, F. Guil, C. Martinez, and R. Marín. Series-Parallel and Tree-Decomposition
Approaches for Fuzzy Constraint Networks. In Proc. of the 8th Ibero-American Conference
on Artificial Intelligence (IBERAMIA’02). LNAI, 2527:275-284, 2002.
3. R. Dechter, I. Meiri, and J. Pearl. Temporal constraint networks. Artificial Intelligence,
49:61-95, 1991.
4. R. Marin, S. Barro, A. Bosch, and J. Mira. Modeling time representation from a fuzzy
perspective. Cybernetic & Systems, 25(2):207-215, 1994.
Parallel State Space Generation and Exploration
on Shared-Memory Architectures
Milan Češka, Bohuslav Křena, Tomáš Vojnar
Faculty of Information Technology

Brno University of Technology
Božetěchova 2, 612 66 Brno, Czech republic
{ceska, krena, vojnar}@fit.vutbr.cz
1 Introduction
Model checking is a technique for an automated formal verification of correctness of
various classes of systems. Compared to the more traditional approaches to this prob-
lem based on simulation and testing, the main advantage of model checking is that the
desirable behavioural properties of the considered system are checked in such a way
that all the reachable states which may affect the given property are guaranteed to be
covered. The disadvantage of model checking is that it may have very high compu-
tational time and space requirements. That is why various ways of dealing with the
computational complexity of model checking are sought.
The systems to be verified may be described using various languages. Here, we
consider the systems to be described by the PNtalk language based on Object-Oriented
Petri Nets (OOPNs). PNtalk [6] has been developed at the Brno University of Technol-
ogy as a tool suitable for modelling, prototyping and rapid development of concurrent
and distributed software systems.
In the paper, we discuss possibilities of parallel OOPN state space generation and
exploration on shared-memory architectures. The goal is to combat the high time com-
plexity of state spaces-based verification methods.
2 Object-Oriented Petri Nets

The OOPN formalism is characterized by a Smalltalk-like object orientation enriched
with concurrency and polymorphic transition execution, which allows for message send-
ing, waiting for and accepting responses, creating new objects, and performing compu-
tations.
OOPNs are based on active objects whose internal activity is described by high-level
Petri nets that are called object nets. The objects communicate via message sending.
Asynchronous reactions of objects to incoming message are described by method nets.
Method nets are also high-level Petri nets and each of them has a set of parameter places
and a return place. Method nets can access places of the corresponding object net. Both
object nets and method nets are dynamically instantiable. Objects can also communicate
synchronously via synchronous ports that resemble special transitions of objects nets,
which can be fired only when they are activated from some classical transition. A simple
OOPN model in Fig. 1 demonstrates modelling by OOPNs on a simple example of
dealing with the stack data structure.
The latest research in the area of OOPNs shows that due to their features, OOPNs
are every suitable for describing open and reflective systems [7].
class name super class object net
class Stack is_a PN
hasDepth: n
synchronous port
parameter x
n
place push: x top pop method net selector
t t
0
x t pop err
NT NT
t>0 t=0 transition guard
push NT:=t+1 NT:=t-1
(NT,x) (t,x)
#ok (#ok,x) #err
PSfrag replacements
method net
output stack
return return
place
Application is_a PN
place with transition action

s1:=Stack new
initial marking t1
p1
s1
output arc
p3 1, 2, 3, 4 p2
input arc
x s1 s1
test arc
t2 s1 push: x y:=s1 pop t3
y
x (#ok,x)
p5 p4
t4
Fig. 1. A simple OOPN model
3 A Parallel OOPN State Space Generator

The main problem of the model checking methods is the so-called state space explo-
sion—the number of states grows exponentially with the size of the model. Thus, it is
difficult or practically impossible to apply these methods directly to large systems. We
can identify two main approaches for dealing with the problem. The first class of these
methods consists in sophisticated generation and exploration of the state spaces while
the second one tries to exploit more powerful computer architectures. In the context of
OOPNs, we have explored the first approach, e.g., in [4, 9, 5]. Here, we consider the
second possibility. Namely, we discuss the techniques behind a parallel state space gen-
erator of OOPNs which we have developed for architectures with shared memory using
the Java programming language and the JOMP tool [2].
The main task in programming parallel applications for shared-memory architec-
tures is to achieve a good load distribution among multiple computation threads taking
into account their need to synchronize when accessing shared data. In our case, we need
to deal with three shared data structures—(1) a hash table, which is used for storing the
state space such that fast searching of states is possible when we need to know whether
a certain state has already been found or not, (2) a list of non-explored states that con-
tains states whose successors have not been explored yet, and (3) a marking pool which
is a special pool for storing decomposed markings.
In order to minimize conflicts among threads trying to access the hash table, we
have divided the table into several parts. In addition, we have divided also the list of
non-explored states (also called as the work list) in such a way that a newly generated
state is put to the unique part of the work list associated with the part of the hash table
to which the state is stored. Moreover, we let the thread that is responsible for handling
a certain part of the hash table use a private work list when processing the states to be
explored. Finally, to further improve the performance, we create a copy of the marking
pool for each thread. The distribution of data we thus obtain is shown in Fig. 2.
Node
Thread
hash table
marking pool
PSfrag replacements
local work list

work list
Fig. 2. Data Distribution
As we have already said, one of the crucial issues for efficiency of parallel applica-
tions is the quality of load sharing among threads. We have started with one data node
(i.e. a fraction of the hash table) for each thread. However, this did not work well. At
first, we are not able to divide the hash table to parts with the same number of states
in advance. Moreover, for a fine load sharing, it is necessary that each thread has some
work (i.e. states to explore) during the whole production run, which is impossible to
predict and guarantee in advance too.
Therefore, we have divided the hash table and the work list to more parts than we
have threads. We find by experiments that the best performance is reached when three
times more nodes than threads are created. Then, threads switch among nodes and try
to find some work to do as it is shown in Fig. 3 for the case of two threads.
Using all these optimizations, we have reached parallel speedups up to 3.3 using 16
processors on the Sun Fire 15k server and speedups up to 2.3 using 4 processors on the
Sun Enterprise 450 server. Moreover, the work distribution among particular threads is
fine (e.g., for 4 threads and about 10 million states: 24.3 % – 25.5 % – 25.3 % – 24.3 %).
The main problem that prevents us to achieve an even better parallel speedup is
the memory management system (especially garbage collecting) due to the memory
operations being performed sequentially. A possible approach to this problem is to im-
plement a user-specific memory management system like in [1], to reduce the number
of dynamically created objects.
N0 N1
Thread 0
N5 N2
PSfrag replacements
Thread 1
N4 N3
Fig. 3. Work Seeking
In the full paper, we describe the proposed algorithm in more detail including a dis-
cussion of efficient ways of storing and comparing the highly complex states of OOPNs,
more experimental results, etc.
Acknowledgement. This work was supported by the Czech Grant Agency under the
contracts 102/04/0780 and 102/03/D211 and it was also supported by the European
Community within the “Access to Research Infrastructure Action of the Improving Hu-
man Potential Programme” under the contract HPRI-CT-1999-00026.
References
1. S. C. Allmaier and G. Horton. Parallel Shared-Memory State-Space Exploration in Stochas-
tic Modelling. In Proc. of IRREGULAR ’97, LNCS 1253, 1997. Springer.
2. M. Bull and M. E. Kambites. JOMP—an OpenMP-like Interface for Java. In Proc. of the
ACM 2000 conference on Java Grande, 2000. ACM Press.
3. E. M. Clarke, O. Grumberg, and D. A. Peled. Model Checking. The MIT Press, Cambridge,
Massachusetts, London, England, 2000.
4. M. Češka, V. Janoušek, and T. Vojnar. Towards Verifying Distributed Systems Using Object-
Oriented Petri Nets. In Proc. of EUROCAST’99, LNCS 1798, 2000. Springer.
5. M. Češka, L. Haša, and T. Vojnar. Partial Order Reduction in Model Checking of Object-
Oriented Petri Nets. In Proc. of EUROCAST’03, LNCS 2809, 2003. Springer.
6. V. Janoušek. Modelling Objects by Petri Nets. PhD. thesis, Brno University of Technology,
Brno, Czech Republic, 1998.
7. B. Kočı́. Methods and Tools for Implementing Open Simulation Systems. PhD. thesis, Brno
University of Technology, Brno, Czech Republic, 2004.
8. B. Křena. Analysis Methods of Object Oriented Petri Nets. PhD. thesis, Brno University of
Technology, Brno, Czech Republic, 2004.
9. T. Vojnar. Towards Formal Analysis and Verification over State Spaces of Object-Oriented
Petri Nets. PhD. thesis, Brno University of Technology, Brno, Czech Republic, 2001.
Tree-structured Legendre multi-wavelets
Ekaterina Pogossova, Karen Egiazarian, Atanas Gotchev, Jaakko Astola
Tampere University of Technology, Signal Processing Laboratory, P. O. Box 553,

FIN-33101 Tampere, FINLAND
Extended Abstract
1 Introduction
Recently, multi-wavelets, that is, wavelets with more than one scaling and wavelet
function have gained a considerable interest in the area of signal processing, in
particular in signal de-noising and compression due to their promising results in
these areas.
The first multi-wavelets were introduced by Alpert [1], who constructed
Legendre type of wavelets on the interval [0,1) with several scaling functions
φ0 , φ1 , . . . , φN −1 , instead of just one scaling function φ0 . This difference enabled
high-order approximation with basis functions supported on non-overlapping in-
tervals. In the particular case of N = 1, Alpert’s multi-wavelets coincide with
the Haar basis.
The discrete analogue of continuous Legendre multi-wavelets has been also
introduced by Alpert [1]. The structure of this analogue is essentially similar
to that of continuous multi-wavelet bases, but the discrete construction is more
convenient when the representation of a function (and its related operations) is
based on its values at a finite set of points [1].
In [3], a new basis, called the tree-structured Haar (TSH) basis was intro-
duced, which is a generalization of the classical Haar basis to the arbitrary time
and scale splitting. The idea behind such a construction was to adapt the basis
to the signal on hand. Discrete orthogonal TSH transform has been successfully
applied to the problem of de-noising signals [5].
In [4], the TSH structure has been extended to the multi-wavelet bases by
the analogy with Alpert’s construction. The construction of the basis functions
assured their orthogonality in the continuous space. However, the question re-
garding the construction of discrete orthogonal bases was left unexplored. In this
paper we address the problem of constructing the discrete counterpart of tree-
structured multi-wavelets, so called, tree-structured Legendre (TSL) transform,
by generalizing the construction procedure, introduced in [1].
2 Continuous Legendre multi-wavelets: non-dyadic

interval splitting
The two-scale relations for the Legendre scaling functions and wavelets of order
N in the case of arbitrary interval splitting point α, introduced in [4], are in the
form: N −1 N −1
φk (x) = j=0 pk,j φj (αx) + j=0 pk,N +j φj ( αx−1
α−1 ),
N −1 N −1 αx−1 (1)
ψk (x) = j=0 qk,j φj (αx) + j=0 qk,N +j φj ( α−1 ).
The parameter α is selected according to the construction of tree-structured
Haar (TSH) basis, introduced in [3]. The name TSH comes from the fact, that
the splitting rules follow from a special tree [3]. In particular, if α = 2, we get
exactly the dyadic Legendre multi-wavelet basis functions. Figure 1 illustrates
the non-symmetric wavelets for N = 2 and α = 3 [4].
3 Discrete counterparts
In [1], an algorithm of constructing discrete Legendre multi-wavelet orthogonal

matrices was introduced. We have generalized the algorithm to the case of arbi-
trarily selected parameter α, as for TSH basis. As a result, discrete counterparts
of tree-structered Legendre (TSL) multi-wavelets have been derived. These new
discrete multi-wavelet like bases can improve the performance for certain signal
processing tasks, such as signal de-noising.
Below we present an algorithm for discrete TSL multi-wavelet construction
that includes Alpert’s algorithm as a special case. We start with constructing a
binary tree, according to which the basis matrix will be derived. Each node of
such a tree has one or two children. Each node is labelled by a number, equal to
the sum of its children’s labels. All terminal nodes are labelled by N , where N
is the required approximation order. Thus, the label at each node is a multiple
of N . An example of such a tree with N = 2 is presented in Figure 2.
By d we denote the depth of the tree, and let µj , j ∈ {0, 1, . . . , d} be the
number of nodes on level j. The tree in Figure 2 is of depth d = 4, with µ0 =
1, µ1 = 2, µ2 = 3, µ3 = 4, and µ4 = 6. Let also c(j, i), i ∈ {1, 2, . . . , µj } be the
nodes on level j, counted from the left to the right, and ν(c(j, i)) be the label at
node c(j, i).
Now we proceed to the definition of basis matrices. Given the set of points
{x1 , x2 , . . . , xn } ∈ R with x1 < x2 < . . . < xn , where n is the label at the root
of the tree, we define so called ”moment matrices”, M1,i , for i ∈ {1, 2, . . . , µd−1 }
2.5 3
2.5
2
2
1.5
1.5
1
1
0.5 0.5
0
0
−0.5
−0.5
−1
−1
−1.5
−1.5 −2
0 0.5 1 0 0.5 1
(a) (b)
Fig. 1. Non-symmetric Legendre multi-wavelets for α = 3, a) wavelet ψ0 (x), b) wavelet

ψ1 (x)
as follows: ⎛ ⎞
1 xui +1 . . . x2N
ui +1
−1
⎜ 1 xui +2 . . . x2N −1 ⎟
⎜ ui +2 ⎟
M1,i =⎜ . . ⎟, (2)
⎝ .. .. ⎠
2N −1
1 xui +2N . . . xui +2N
if c(d − 1, i) is a splitting node, and
⎛ ⎞
1 xui +1 . . . x2N
ui +1
−1
⎜ 1 xui +2 . . . xu +2
2N −1 ⎟
⎜ i ⎟
M1,i = ⎜ . .. ⎟, (3)
⎝ .. . ⎠
1 xui +N . . . x2N −1
ui +N
i−1
if c(d − 1, i) is a non-splitting node. Here ui = j=1 ν(c(d − 1, j)) for i >
1, and u1 = 0. Additionally, we define matrices U1,i = (Orth(M1,i ))T , i ∈
{1, 2, . . . , µd−1 }, where Orth stands for column-by-column Gram-Shmidt orthog-
onalization.
By S U and S L let us denote the upper and lower k rows of (2k×2k) matrix S,
respectively. The first (n × n) basis matrix, U1 , is obtained by first constructing
an auxiliary matrix, Ũ1 :
⎛ L ⎞
Ũ1,1
⎜ L
Ũ1,2 ⎟
⎜ ⎟
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ L ⎟
⎜ Ũ1,µd−1 ⎟
U˜1 = ⎜ U ⎟, (4)
⎜ Ũ1,1 ⎟
⎜ ⎟
⎜ U
Ũ1,2 ⎟
⎜ ⎟
⎜ .. ⎟
⎝ . ⎠
U
Ũ1,µ d−1
where ⎧
, if c(d − 1, i) is a splitting node
⎨ U1,i
Ũ1,i = IN (5)
⎩ , otherwise.
0N
12
8 4
4 4 4
4 2 2 4
2 2 2 2 2 2
Fig. 2. TSH Legendre tree, N = 2

In (5) IN and 0N denote (N × N ) identity and zero matrices, respectively. U1 is
now obtained by deleting all zero rows from Ũ1 .
The m − th basis matrix, m ∈ {2, . . . , d}, is Um × Um−1 × · · · × U1 , with
(n × n) matrices Uj , j ∈ {2, . . . , k} defined by the formula

Uj
Uj = . (6)
I
Here Uj are obtained by deleting zero rows from Ũj :
⎛ L ⎞
Ũj,1
⎜ ⎟
⎜ ⎟
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ L ⎟
⎜ Ũj,µ ⎟
U˜j = ⎜ U d−j
⎟, (7)
⎜ Ũj,1 ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ .. ⎟
⎝ . ⎠
U
Ũj,µd−j
Ũj,i and Uj,i are related in the same way as Ũ1,i and U1,i (see (5)), Uj,i =
(Orth(Mj,i ))T , and Mj,i is given by

U
Uj−1,k × Mj−1,k
Mj,i = (8)
U
Uj−1,k+1 × Mj−1,k+1
if c(d − j, i) is a splitting node and c(d − j + 1, k), c(d − j + 1, k + 1) are its
left and right children, respectively; and Mj,i = Mj−1,k if the node c(d − j, i) is
non-splitting, and c(d − j + 1, k) is its only child.
The final matrix, U = Ud × Ud−1 × · · · × U1 , represents the discrete wavelet-
like basis of parameter N on x1 , x2 . . . , xn . As it has been mentioned in [1], some
adjustments must be made to this formula to ensure numerical stability. These
issues are discussed in [2] and we will not cover them here.
References
1. Alpert, B.: A class of bases in L2 for the sparse representation of integral operators,
SIAM J. Math. Anal., 24, (1993), 246–262
2. Alpert, B., Beylkin, G., Coifman, R., Rochlin, V.: Wavelet bases for the fast solution
of second-kind integral equations, SIAM Journal on Scientific Computing, 14, 1,
1993, 159–184
3. Egiazarian K., Astola, J.: Tree-structured Haar transforms, Journal of Mathematical
and Imaging Vision, 16, 2002, 267–277
4. Gotchev, A., Egiazarian, K., Astola, J.: On tree-structured legendre multi-wavelet
bases, Proc. Int. Workshop on Spectral Methods and Multirate Signal Processing
(SMMSP 2001), June 16-18, Pula, Croatia, 2001, 51–56
5. Pogossova , E., Egiazarian, K., Astola, J.: Signal de-noising in tree-structured Haar
basis, Proc. 3rd International Symposium on Image and Signal Processing and Anal-
ysis (ISPA’03), Rome, Italy, September 18-20, 2, 2003, 736–739
Remarks on Calculation of Autocorrelation on
Finite Dyadic Groups by Local Transformations
of Decision Diagrams
Radomir S. Stanković1 and Mark G. Karpovsky2

1
Dept. of Computer Science, Faculty of Electronics
Beogradska 14, 18 000 Niš, Serbia
2
Dept. of Electrical and Computer Engineering, Boston University
8 Saint Marry’s Street, Boston Ma 02215, USA
1 Introduction
Autocorrelation is an important operation in signal processing and systems the-
ory [1]. In particular, the autocorrelation on finite dyadic groups, denoted as
dyadic autocorrelation Bf , (see Definition 1) is useful in switching theory and
design of systems whose inputs and outputs are represented by functions defined
in 2n , n ∈ N points, including switching functions as an example [2], [3], [6],
[7], [8], [9]. Recently, some new applications of dyadic autocorrelation in spectral
methods for switching functions, testing of logic networks [4] and optimization
of decision diagrams (DDs) for representation of discrete functions has been re-
ported [10]. In this paper, we define and discuss a method for calculation of
the dyadic autocorrelation through decision diagrams, the use of which permits
processing of functions of a large number of variables.
2 Background Theory
Denote by C2n the finite dyadic group, where C2n = ×ni=1 C2 , and C2 = ({0, 1}, ⊕),
where ⊕ denotes multiplication modulo 2 (EXOR).
Definition 1 For a function f : C2n → R, where R is the field of real numbers,
P2n −1
the autocorrelation Bf is defined by Bf (τ ) = x=0 f (x)f (x ⊕ τ ), where τ =
0, . . . , 2n − 1. In binary notation, x = (x1 , . . . , xn ) and τ = (τ1 , . . . , τn ), where
xi , τi ∈ {0, 1}.
In matrix notation, if f and Bf are represented by vectors F = [f (0), . . . , f (2n −
1)]T and Bf = [Bf (0), . . . Bf (2n − 1)]T , respectively, then, Bf = B(n)F, where
B(n) is the dyadic autocorrelation matrix for f .
2.1 Wiener-Khinchin theorem

The Walsh transform is defined by the Walsh matrix
n
O
W(n) = W(1),
i=1
·¸
1 1
where ⊗ denotes the Kronecker product, and W(1) = .
1 −1
The Wiener-Khinchin theorem states a relationship between the autocorre-
lation function and Walsh (Fourier) coefficients [2]
Bf = 2n W −1 (W f )2 ,
where W denotes the Walsh transform operator.
3 Decision Diagrams
Decision diagrams are data structures providing compact representations of dis-
crete functions defined in a large number of points [11]. In this paper, we as-
sume that a given function f is represented by a Multi-terminal Binary DD
(MTBDD(f )) [11]. A MTBDD is a directed acyclic graph consisting of non-
terminal nodes and constant nodes connected by edges. Each node has two out-
going edges labeled by the negative and positive literals xi and xi of the decision
variable assigned to the node. Nodes to which the same variable is assigned form
a level in the MTBDD.
Notice that in calculations over decision diagrams, the impact of the deleted
nodes should be taken into account through the cross points defined as points of
intersections of paths from the root node to the constant nodes with the imag-
inary lines showing levels in decision diagrams, which means lines connecting
nodes to which the same decision variable is assigned [12]. Complexity of a de-
cision diagram is usually expressed in terms of the number of non-terminal and
constant nodes, called the size of the decision diagram.
f f Bf
_ S2 _ S2 _ 2S
x1 x1 x1 x x1 x1
1
S2 S
_ S2 _
S2
_ 2
x2 x2 x2 x2 _ x2 x2 x2
x2
_ S2 _ S2 _ S2 _ S2 S2 _ S2
x3 x3 x3 x3 x3 x3 x3 x3 x3 x3
0 0 1 2 3 3 3 3 0 1 2 3 41 40 36 18
Fig. 1. MTBDT(f ), MTBDD(f ) and MTBDD(Bf in Example 1.
Example 1 Fig. 1 shows a MTBDT and the corresponding MTBDD for a func-
tion f given by the vector of function values F = [0, 0, 1, 2, 3, 3, 3, 3]T . In this
figure, we also show the cross points in the MTBDD. Fig. 1 shows the MTBDD
for the autocorrelation function Bf (τ ).
4 Wiener-Khinchin Theorem over Decision Diagrams

The Walsh spectrum Sf of a given function f represented by a MTBDD is deter-
mined by performing at each node and the cross point of the MTBDD(f ) the cal-
culations determined by W(1). For simplicity, we say the nodes and cross points
in MTBDD(f ) are processed by W(1). In this way, MTBDD(f ) is converted
into the MTBDD(Sf ). We perform the multiplication of Sf by itself by replac-
ing the values of constant nodes Sf (i) with Sf2 (i) [11]. Then, the MTBDD(Bf )
is determined by performing the calculations determined by W(1) at each node
and the cross point of the resulting MTBDD(Sf ) followed by the normalization
with 2n , since the Walsh matrix is self-inverse up to the constant 2−n .
4.1 Complexity of the method

Since in calculation of the Walsh spectrum, we perform an addition and a sub-
traction at each node and the cross point distributed over n levels, the com-
plexity is O(2n · size(M T BDD(f ))). Notice that the number of cross points
in a MTBDD is on the average at about 30% of the number of non-terminal
nodes [11]. The result of these calculations is the MTBDD(Sf ). Then, we per-
form squaring of the values of constant nodes and perform the inverse transform.
Thus, since the Walsh transform is self inverse, the complexity of these calcula-
tions is O(2n · size(M T BDD(Sf ))). After multiplication with the scaling factor
2n , the MTBDD(Bf ) is derived.
Notice that the size of the MTBDD for the Walsh spectrum is usually greater
than that of the MTBDD for functions with a limited number of different val-
ues. Since in calculation of the autcorrelation function, MTBDD(f ) is converted
into a MTBDD(Sf ), which is in many cases larger in terms of size than the
MTBDD(f ), the space complexity of the method is O(size(M T BDD(Sf ))).
5 In-place Calculation of Autocorrelation Coefficients

We define a transformation of nodes in MTBDDs that consists of permutation
of labels at the outgoing edges.
The i-th row of the autocorrelation matrix is the vector of function values
f (x ⊕ i), where ⊕ denotes the componentwise EXOR over the binary represen-
tations for x = (x1 , . . . , xn ), and i = (i1 , . . . , in ). In decision diagrams, this shift
of the argument for f implies permutation of labels at the edges of some nodes
in the decision diagram for f . Nodes whose edges should be permuted are situ-
ated at the levels whose position within the decision diagram corresponds to the
position of 1-bits in the binary representation for the row index i.
The i-th autocorrelation coefficient is calculated by the multiplication of the
i-th row of the autocorrelation matrix Bf by the vector F of function values
for f . When f and rows of Bf are represented by decision diagrams, it follows
that the i-th autocorrelation coefficient is calculated by the multiplication of the
decision diagrams for f and f (x ⊕ i). This can be performed by the classical pro-
cedure for multiplication of decision diagrams. However, since decision diagrams
for f (x) and f (x ⊕ i) differ in labels at the edges, in practical programming
implementations, calculations can be organized over a single diagram similar
as calculations of FFT can be organized in-place [1]. Therefore, complexity of
calculation is proportional to the number of nodes in the decision diagram for
f.
In this procedure, f is represented by a MTBDD which is then traversed in
such a way to multiply values of constant nodes in the MTBDD for f with the
values of constant nodes in the MTBDD for f (x ⊕ τ ) and perform the addition
of these values. The way of traversing is determined by the binary components
τi , i = 0, 1, . . . , n − 1 of τ . A flag is associate to each non-terminal node, to
show if the node was already traversed. In this manner, the coefficient Bf (τ ) is
calculated. The procedure has to be repeated for each coefficient.
References
1. Agaian, S., Astola, J., Egiazarian, K., Binary Polynomial Transforms and Nonlin-
ear Digital Filters, Marcel Dekker (1995).
2. Karpovsky, M.G., Finite Orthogonal Series in the Design of Digital Devices, John
Wiley (1976).
3. Karpovsky, M.G., (ed.), Spectral Techniques and Fault Detection, Academic Press
(1985) 35-90.
4. Karpovsky, M.G., Stanković, R.S., Astola, J.T., Spectral techniques for design and
testing of computer hardware”, Proc. Int. Worksop on Spectral Techniques in Logic
Design, SPECLOG-2000, Tampere, Finland, June 2-3 (2000) 1-34.
5. Meinel, Ch., Somenzi, F., Tehobald, T., Linear shifting of decision diagrams and
its application in synthesis, IEEE Trans. CAD, Vol. 19, No. 5 (2000) 521-533.
6. Pichler, F., Walsh functions and linear system theory, Proc. Applic. Walsh Func-
tions, Washington, D.C. (1970) 175-182.
7. Pichler, F., Realizations of Prigogine’s Λ-transform by dyadic convolution, in
Trappl, R., Horn, W., (eds.), Austrian Society for Cybernetic Studies, ISBN
385206127X (1992).
8. Rice, J.E., Muzio, J.C., Methods for calculating autocorrelation coefficients, Proc.
4th Int. Workshop on Boolean Problems, (IWSBP2000) (2000) 6976.
9. Rice, J.E., Muzio, J.C., Use of autocorrelation function in the classification of
switching functions, Euromicro Symposium on Digital System Design (2002) 244-
251.
10. Rice, J., Serra, M., Muzio, J.C., The use of autocorrelation coefficients for variable
ordering for ROBDDs, Proc. 4th Int. Workshop on Applications of Reed-Muller
Expansion in Circuit Design, Victoria, Canada, August 20-21 (1999) 185-196.
11. Sasao, T., Fujita, M., (eds.), Representations of Discrete Functions, Kluwer Acad-
emic Publishers (1996).
12. Stanković, R.S., Sasao, T., Moraga, C., Spectral transform decision diagrams, in T.
Sasao, M. Fujita, (ed.), Representations of Discrete Functions, Kluwer Academic
Publishers (1996) 55-92.
A new Pseudo-Random Generator based on Gollmann
Cascades of Baker-Register-Machines
Dominik Jochinger, and Franz Pichler
Systems Theory, Johannes Kepler University Linz

Altenberger Str. 69, A-4040 Linz, Austria
Extended Abstract
In this paper, we present a new pseudo-random sequence generator, constructed by

the generalized discrete Baker transformation. This new generator is called Cascaded
Baker Register Machine (CBRM), which combines chaotic behaviour with the
strength of the shift-register theory.
It is shown that a CBRM has good properties of randomness, such as large periods
and high linear complexity. It can provide high security with rather fast encryption
speed, and can be realized simply by both hardware and software.
The chaotic function that we use is the well known generalized discrete Baker
transformation [2]. The chaotic properties of the Baker transformations have been
investigated by [3].A CBRM is similar well known Gollmann cascade [1, 4] of k
stages. It consists of a series of k Baker-Register-Machines (BRM), of same length
c
111…11
clock clock clock
….
…. …. ….
Fig. 1. Example of a Cascaded Baker Register Machine of 3 stages
The partition-key π which defines a specific generalized Baker Transform can be

customer specified.It is used to initialize a Baker-Register-Machine. The key π
determines the cyclic length of the BRM. Not every possible partition-key leads to a
secure cyclic length. The table below shows some possible partitions of a 16 × 16
Baker map and their cyclic length.
Table 1. Some possible partitions

partition cyclic length
(8,4,4) 622440
(4,4,8) 622440
(4,4,4,2,2) 1680
(4,4,2,2,4) 55440
(4,2,2,4,4) 55440
(2,2,4,2,2,2,2) 55440
(2,2,2,2,4,2,2) 55440
An analysis of the CBRM shows that it has good statistical properties. For example
n-gram distributions are uniform.Another nice property of a BRM is that the
generated sequences are highly sensitive to the initial key. By statistical analysis, we
found that changing one bit in the encryption key caused 49.4% of the bits to change
in the corresponding cipher-image. Optimally, fully half of the bits should change, in
order to hide any information about the key from leaking.The linear complexity
profile of the CBRM algorithm is staying very close to the length of the generated
sequence.The cycle length (period) of the CBRM depends on the partition key, as
mentioned above. In general it is difficult to compute the exact cycle length of the
CBRM. A general formula for this computation can be given as follows:
Let r be the number of the feedback-shift-registers that realize the Baker-Register-
Machine, σ i (i =1, 2, …, r) the length of the shift-register i, and k the number of
cascaded stages. Then the cycle length of the CBRM is given by:
L = lcm (σ 1 , σ 2 ,...., σ r )k
As an example L for n = 256, k = 10, π ={4, 4, 8}. We get L = 62244010

≈ 8.7 ⋅ 10 57 . Such a length is enough for many practical applications.
The CBRM has been simulated and tested in VHDL (hardware) and JAVA™
(software).
To evaluate the possibility for an application of the Baker-transformation in stream
ciphering further cryptological research is needed.
References
1 Dieter Gollmann, „Kaskadenschaltungen taktgesteuerter Schieberegister als

Pseudozufallszahlengeneratoren“, PhD thesis, Johannes Kepler Universität Linz, VWGÖ-
Verlag, Wien, 1986.
2 Franz Pichler, Josef Scharinger, “Ciphering by Bernoulli-Shifts in Finite Abelian Groups”,
Contributions to General Algebra (ed. G. Pilz), Hölder-Pichler-Tempsky, pp. 249-256,
Wien, 1995.
3 Josef Scharinger, „Experimentelle harmonische Anaylse von Bäcker-dynamischen 2D
Systemen und ihre Anwendung in der Kryptographie“, Dissertation, Uni-Linz, 1994.
4 Bruce Schneier, „Applied Cryptography, Protocols, Algorithms and Source Code in C,
Second Edition“, John Wiley and Sons, pp. 445-446, 1995.
An Excellent Permutation Operator for
Cryptographic Applications
Josef Scharinger
Johannes Kepler University, Institute of Computational Perception,

4040 Linz, Austria
Josef.Scharinger@jku.at
1 Introduction
Permutations are a core component of almost every cipher. No matter if we
consider the DES [5], AES [2] or most of the other encryption algorithms [6]
relevant nowadays, we always find permutation operators as essential building
blocks inside. In this contribution we will introduce key-dependent permutation
operators of provably excellent quality inspired by chaotic Kolmogorov flows [1,
4]. From chaotic systems theory it is known [3] that the class of Kolmogorov flows
exhibits the highest degree of instability among all dynamical systems. As will
be demonstrated in the sequel, these outstanding properties make them a per-
fect inspiration for developing a novel class of strong cryptographic permutation
operators.
2 Chaotic Kolmogorov Flows

Continuous Kolmogorov flows (see e.g. Figure 1) perform iterated stretching,
squeezing and folding upon the unit square E. It is guaranteed [1, 3] that their
application leads to perfect mixing of the elements within the state space.
Formally this process of stretching, squeezing and folding Pk is specified as fol-
lows. Given a partition π = (p1 , p2 , . . . , pk ) (0 < pi < 1, i=1 pi = 1) of the unit
interval U and stretching and squeezing factors defined by qi = p1i . Furthermore,
let Fi defined by F1 = 0 and Fi = Fi−1 + pi−1 denote the left border of the
vertical strip containing the point (x, y) ∈ E to transform. Then the continuous
Kolmogorov flow Tπ will move (x, y) ∈ [Fi , Fi + pi ) × [0, 1) to the position
y
Tπ (x, y) = (qi (x − Fi ), + Fi ). (1)
qi
3 Discrete Kolmogorov Systems

Chaotic Kolmogorov flows perform perfect mixing of the continuous unit square
but how can we utilize them to mix a discrete data block of dimensions n×n? We
have developed the following formula. Given a list δ = (n1 , n2 , . . . , nk ) (0 < ni <
Pk
n, i=1 ni = n) of positive integers that adhere to the restriction that all ni ∈ δ
Fig. 1. Illustrating the chaotic and mixing dynamics associated when iterating a Kol-
mogorov system.
must partition the side length n. Furthermore let the quantities qi be defined by
qi = nni and let Ni specified by N1 = 0 and Ni = Ni−1 + ni−1 denote the left
border of the vertical strip that contains the point (x, y) to transform. Then the
discrete Kolmogorov system Tn,δ will move the point (x, y) ∈ [Ni , Ni +ni )×[0, n)
to the position
Tn,δ (x, y) = (qi (x − Ni ) + (y mod qi ), (y div qi ) + Ni ). (2)
In the final paper we will prove the following theorem:
Theorem 1. Let the side-length n = pm be given as integral power of a prime
p. Then the discrete Kolmogorov system Tn,δr as defined in equation 2 fulfills
the properties of ergodicity, exponential divergence and mixing provided that at
least 4m iterations are performed and lists δr used in every step r are chosen
independently and at random.
Based on this theorem it is well justified to claim that the permutation op-
erator developed in this contribution is indeed an excellent key-dependent per-
mutation operator for cryptographic applications.
References
1. V.I. Arnold and A. Avez. Ergodic Problems of Classical Mechanics. W.A. Benjamin,
New York, 1968.
2. J. Daemen and V. Rijmen. AES proposal: Rijndael. In First Advanced Encryption
Standard (AES) Conference, Ventura, CA, 1998.
3. S. Goldstein, B. Misra, and M. Courbage. On intrinsic randomness of dynamical
systems. Journal of Statistical Physics, 25(1):111–126, 1981.
4. Jürgen Moser. Stable and Random Motions in Dynamical Systems. Princeton Uni-
versity Press, Princeton, 1973.
5. US Department of Commerce National Bureau of Standards. Data encryption stan-
dard. NBS FIPS PUB 46-1, 1977.
6. Bruce Schneier. Applied Cryptography. Addison-Wesley, 1996.
Fault Cryptanalysis of ElGamal and Related
Signature Schemes
Janusz Biernat and Maciej Nikodem

Abstract. In this article the immunity of ElGamal and related signa-

ture schemes against fault cryptanalysis (FA) is examined. Although such
schemes have been widely adopted, their resistance against fault crypt-
analysis has not been verified in detail. Majority of those schemes are
not immune to fault cryptanalysis and can be broken without solving
discrete logarithm problem. It will be proved that the selected signature
schemes can be broken in O(n2 logn) steps if single bit-flip errors are in-
ducted during computations. We will also present methods that allow to
improve security of ElGamal and DSA.
1 Introduction
In this paper we focus on six different versions of ElGamal and related DSA
and Schnorr signature schemes. It will be shown that, choosing algorithm to a
particular application has huge impact to security, which will decrease when one
is able to induct errors during computations. Our attack assumes that signing
purposes are implemented in tamper-proof device and one is able to induct
random bit-flip error to private key a stored in this device. We also assume
that each error affects only one bit and the probability that i-th bit of key a is
erroneous equals 1/n.
In our paper we present different signature schemes and their weakness
against fault analysis. We also present methods that improve some of those
schemes to remain secure. Those observations can be useful when modifying or
creating new signature schemes.
2 ElGamal Signature Scheme

In all versions of ElGamal algorithm public and private keys are defined as
follows:
– private key - a
– public key - (p, g, y)
where p is a large prime, g is a generator of the multiplicative group Zp∗ and y is
calculated as g a modp. Signature generation, in case of the most popular version
of ElGamal scheme, goes as follows:
1. select a random integer k < p − 2 with gcd(k, p − 1) = 1
2. compute r = g k modp
3. compute s = k −1 (h(m) − ar)mod(p − 1) where h(m) is a hash function
and then the signature for m is a pair (k, s)
The scheme mentioned above is susceptible to fault cryptanalysis and one can
recover the key a with probability 1/2 after nlog2n erroneous signings. Complex-
ity of this attack is O(n2 logn). This can be easily proved.
Similar analysis can be done with respect to other versions of ElGamal
schemes. Those versions differs only in step 3 of signing which can have one
of the following forms:
3a. s = a−1 (h(m) − kr)
3b. s = ar + kh(m)
3c. s = ah(m) + kr
3d. s = a−1 (r − kh(m))
3e. s = k −1 (r − ah(m))
As we show in our paper versions 1 and 4 are immune to fault cryptanalysis
of the discussed type. On the other hand other versions of ElGamal scheme
are susceptible to fault cryptanalysis and can be broken with probability 1/2 in
nlog2n attempts. We will also present and examine modifications that allow to
increase cryptanalysis complexity and thus improve security.
3 DSA and Schnorr Schemes
In our paper we also study DSA and Schnorr signature schemes. We show that
DSA scheme cannot be broken using the same technique as in case of ElGammal
thanks to mathematical modifications that influence key preparation and signing
processes. Unfortunately there are other methods how one can break DSA. Those
methods were presented in [2].
We have also examined Schnorr signing scheme. In the article we present
how one can recover secret integer a from tamper-proof device without solving
discreet logarithm problem or reversing one-way function. This can be done
despite the Schnorr scheme use the same mathematical backgrounds for key
generation as DSA scheme. Moreover such attack requires less mathematical
computations and can be carried out easier than in case of ElGamal scheme.
References
1. Boneh D., DeMillo R.A., Lipton R.J., On the Importance of Checking Crypto-
graphic Protocols for Faults, Advances in Cryptology - EUROCRYPT’97, LNCS
1233, pp.37-51, Springer-Verlag, 1997
2. Christophe Giraud and Erik W. Knudsen, Fault Attacks on Signature Schemes,
ACISP 2004, LNCS 3108, pp. 478491, Springer-Verlag, 2004.
Complexity-Theoretical Approaches to
the Design and Analysis of
Cryptographical Boolean Functions
Juan David González Cobas and José Antonio López Brugos
Departamento de Informática
Universidad de Oviedo
{cobas,brugos}@epsig.uniovi.es
1 Introduction
Boolean functions are at the core of the design of every symmetric cipher algo-
rithm, be it stream or block type. There is a wide body of theory supporting the
design choices behind many of the standard algorithms in wide use [4].
However, in many respects the design of cryptographic-quality boolean func-
tions still depends on many ad-hoc constraints. Efficiency and compactness con-
siderations leave very little room for choice when other algorithm parameters
(key schedule, number of rounds, etc.) determine too stringently the properties
we need for the required S-boxes or feedback functions. Choosing at random is
possible if efficiency is neglected [1].
Although complexity theory lies deeply at the roots of cryptographic strength,
it seems too often neglected in the design and analysis of algorithms, in which
algebraic approaches are favored. However, measures of complexity have proven
useful for detecting weaknesses in preliminary phases of design [3].
In this paper we survey the usual criteria that apply to the design of boolean
functions for stream and block ciphers, proposing a complexity-theoretic frame-
work in which measures of complexity provide a test bench for nonlinear elements
of cryptographic algorithms in preliminary analysis.
2 Usual requirements on boolean functions

Established criteria for boolean function quality of design can be enumerated as
follows:
– being balanced
– propagation criteria of various orders
– being bent
– avalanche criteria
The relation among all these criteria and many others have been thoroughly
studied and still are subject of intensive research [6, 5]. Their algebraic character
is remarkable.
3 Complexity-theoretic approaches
With a complexity-theoretic point of view, complexity measures can be applied
to test how good a non-linear boolean function could perform. Complexity mea-
sures that have been proposed, with and without cryptographic purposes in
mind, have been reviewed in [2]. Application of these measures can be done in
several ways:
1. Profiling complexity of pseudo-random sequence generators.

2. Measuring computational complexity of well-known attacks (be it ciphertext-
only, known-plaintext or chosen-ciphertext attacks).
3. Computing the number of rounds needed to achieve some complexity lower
bound to assess the quality of nonlinear elements.
4. Boolean circuit complexity evaluation of selected parts of a network.
5. Estimating the Kolmogorov complexity of generic cryptanalysis strategies
applied to subalgorithms and/or weakened forms of full algorithms.
4 A test bench
In this presentation we propose 4 and 5, together with the other well-established
techniques as procedures for assessing the suitability of boolean functions for
a given cryptographic purpose. In the framework of a reduced-round version of
the well-known block algorithms AES and Serpent, and a toy, proof-of-concept
stream cipher algorithm, we show how to apply this method to select the best
possible choice for the non-linear elements involved.
References
1. C. M. Adams and S. E. Tavares. Good s-boxes are easy to find. In Proceedings
of the 9th Annual International Cryptology Conference on Advances in Cryptology,
pages 612–615. Springer-Verlag, 1990.
2. J. D. Gonzalez-Cobas and J. A. Lopez-Brugos. A complexity-theoretic approach to
the design of good measures of cryptographic strength. In International Workshop
on Computer Aided Systems Theory, LNCS, volume 8, 2001.
3. J. L. Massey. Shift-register synthesis and BCH decoding. IEEE Transactions on
Information Theory, IT-15:122–127, 1969.
4. W. Meier and O. Staffelbach. Nonlinearity criteria for cryptographic functions. In
Proceedings of the workshop on the theory and application of cryptographic tech-
niques on Advances in cryptology, pages 549–562. Springer-Verlag New York, Inc.,
1990.
5. J. Seberry, X.-M. Zhang, and Y. Zheng. Structures of cryptographic functions
with strong avalanche characteristics (extended abstract). In Proceedings of the
4th International Conference on the Theory and Applications of Cryptology, pages
119–132. Springer-Verlag, 1995.
6. Y. Zheng and X.-M. Zhang. On relationships among avalanche, nonlinearity and
correlation immunity.

! " # $
" """ "
% &" " # $' Æ
" " (
)
*+, "
Æ - . . # $
"
) #$
" " ) / #/$' '
" ) " Æ
"' " "
" ! *0, ) *1,

" "
" *2, !
3
"' 4 "
""' ""' " '
*5, ( '
" )
) 6" 3'
3 "
4 "
"' *7,' Æ "
' 8 9 #
$ '
/
"
"
#
$ ! "
" 8 )
" " % "'

¼

!
" "
#$
%&&'(&')(&'(&
" 8 """
" " 9 # $
:
5
9 5
1

9
5
1

'
3
) "" ' # $
/ ; '
" # 9 5
1

$'
# 9 5
1

$ "
# 9 5
1

$' "

<
9 #

$
9 5
1

'
"
9 5
1

' 9

"

' "" " "

< 9 ## $ " $
# 9 5
1

9 5
1

$

/ " ' 9 :
" =
/<
9 :'

' 9 # $

9 5
1

' '
9 :' "
#95'1' '$' / )
4 "

5 #>
5
1

9

9 9 5$
9 5' / )
# $ 9
5
1

< #

5
1

# $ 9

$ 9 #

$ "

*
" + , +- Æ
.
(/0!
"
# ! *
# !
% #1!
* 2
3&&4 56(& 2)6
7

7 8 * 8- .
(90!
"
"
:! (
!34 354 (5)
79! ; , -

** **!
*"
!
:! !34 3&&&4 5(&
' %!"0
;! 9< -
90!
"
*!
=

(
* 8;( 35)4 5(&'
) %!"
;! ,"
8- >0
!! 2#(
*
?
(
90!
"

3564 () 26
6 #
"

#- 2(

" * (

@
A1!

3554 5('& 2)6

½ ¾
½

¾

!"#
$$ %%& '

! !
! " !
# ! " $ ! " %
& ' ( )* !
$#$# +
,-)% !

. " !
! !
" + " !
" #" "
! / "
" " "
! !
+ 0
! " "
"
! ! / 123 !
# ! "
0 $
/ " &
' 0 " 0 "
" /
+
"
/ " 0
/
!
#

¼
(

#"

!)
# *%%$+%$,+%$+%
+ ! "
! ! 143 ,
" 5
163 "
"
!
!
! " # !
Æ 7 .

8

99 #"
/

0 " " 0 "8
¯
8

" 0
¯
8

" 0
¯ 8

" 0
#
0

)
" & ' 0
" ! ,
! !

*# - . (
#
#

/ "
!0#
1 ' "

/ 2
/
&$ 34 3554 %$
*# - !
# . *

6

#7 8 " !00 #

#
# 15 9 &,
3554 $+$&
!
# :;
; (. "

+ #

!#
%% < 3%%4 ,+,&

Protocol Analysis for Concrete Environments
Dieter Gollmann
TU Hamburg-Harburg
diego@tu-harburg.de
Extended Abstract
In protocol analysis we are given a description of the protocol, a set of security

goals, and a model of the underlying communications system. We then check
whether the protocol meets its desired goals. The research community has dis-
cussed conventions for describing protocols, and there exist proposals for protocol
description languages such as CAPSL1 . Equally, the importance of distinguish-
ing security properties in all their nuances is widely acknowledged and there
has been much research on this topic. Current definitions for the properties of
cryptographic primitives and mechanisms can be found in [6]. A discussion of
various authentication properties is given in [4], to give just two examples.
In contrast, rather less effort has been spent on modelling the communications
environment. Indeed, protocol analysis often tries to assume as little as possible
about the communications system and gives all messages to the adversary for
delivery. This approach is often presented as analysis in the Dolev-Yao model
[3]. The model makes two independent assumptions:
– Cryptography is “perfect”. The adversary does not try to exploit any weak-
ness in the underlying cryptographic algorithms but only algebraic properties
of cryptographic operators and interactions between protocol messages.
– The adversary can observe and manipulate all messages exchanged in a pro-
tocol run and can itself start protocol runs.
The second assumption was already stated by Needham and Schroeder [7]:
We assume that the intruder can interpose a computer in all communica-

tion paths, and thus can alter or copy parts of messages, replay messages,
or emit false material. While this may seem an extreme view, it is the
only safe one when designing authentication protocols.
Analysing protocols in a setting as general as possible is, however, not nec-

essarily a route to higher security. Protocols may make use of features of the
particular environment they were designed for so showing that a protocol does
not meet is goal in a more general setting is useful side-information but should
not be automatically classified as an attack.
1
For a CAPSL tutorial see http://www.csl.sri.com/users/millen/capsl/
We can analyze protocols that should meet well established security require-
ments and use established security primitives. On the other hand, we might
analyze protocols that should meet novel requirements. In the first case, security
goals, assumptions about the environment, and standard cryptographic primi-
tives can be integral parts of the methodology, so that changes to any of these
aspects would require some redesign of the methodology.
In the second case, we need agile methodologies where specific adversaries
(rather than the general Dolev-Yao adversary) and new security requirements
can be defined conveniently. Note that in most cases new protocols are designed
because new requirements have emerged, so that traditional security assumptions
have to be adjusted. For illustration, we briefly sketch four specific scenarios,
together with observations on how established security assumptions may change.
In the binding update protocol for mobile IPv6 [2, 5] a correspondent node
uses return routability to verify a binding update from a mobile node. Although
the mobile node computes a MAC on the binding update confirmation, the keys
(or should we treat those keys better as nonces?) are sent in the clear.
Protocols for the “real” Internet have to consider so-called middleboxes such
as Network Address Translators (NATs) and firewalls. A simple “Alice & Bob”
model of communications is no longer appropriate when designing, and analysing
security protocols at the IP layer.
Multi-layered protocols such as the variants of EAP [1] use identifiers at
different layers. Security analysis has thus to check the binding that is intended
(or not intended, when privacy is a goal) between the different identifiers of a
single principal. Again, the Alice & Bob view of the world is too simplistic.
A protocol providing data integrity in a sensor network described in [8]
achieves its stated goal not by relying on independent witnesses but does not
provide data origin authentication at the same time.
In conclusion, we argue that protocol analysis techniques have to be able to
capture specific aspects of the communications system to be able to properly
analyze the new security protocols that are being defined today.
References
1. B. Aboba, L. Blunk, J. Vollbrecht, J. Carlson, and H. Levkowetz. Extensible Au-
thentication Protocol (EAP), June 2004. RFC 3775.
2. Tuomas Aura, Michael Roe, and Jari Arkko. Security of Internet location manage-
ment. In Proceedings of the 18th Annual Computer Security Applications Confer-
ence, pages 78–87, December 2002.
3. Danny Dolev and Andrew C. Yao. On the security of public key protocols. IEEE
Transactions on Information Theory, IT-29(2):198–208, March 1983.
4. Dieter Gollmann. Authentication by correspondence. IEEE Journal on Selected
Areas in Communications, 21(1):88–95, January 2003.
5. D. Johnson, C. Perkins, and J. Arkko. Mobility Support in IPv6, June 2004. RFC
3775.
6. Alfred J. Menezes, Paul C. van Oorschot, and Scott A. Vanstone. Handbook of
Applied Cryptography. CRC Press, Boca Raton, FA, 1997.
7. Roger M. Needham and Michael D. Schroeder. Using encryption for authentication
in large networks of computers. Communications of the ACM, 21:993–999, 1978.
8. Harald Vogt. Integrity preservation for communication in sensor networks. Technical
Report 434, ETH Zürich, February 2004.
Pattern Recognition in AVHRR Images by Means of
Hibryd and Neuro-Fuzzy Systems
Jose Antonio Piedra, Francisco Guindos, Alberto Molina, and Manuel Canton
Universidad de Almeria, Departamento de Lenguajes y Computacion, Phone +34 950 0140

36, 04120 Almeria, Spain
{jpiedra, fguindos}@ual.es, alberto@altoalmanzora.org
mcanton@ual.es
Extended Abstract
The problem is to study methods for the automatic interpretation of ocean satellite
images to recognize meso and macroscalar ocean structures. The difficulty of the
image analysis and understanding problem for satellite data is due, in large part, to the
lack of a precise mathematical description of the observed structures and to their
variability. This work proposes an approximation that show the results obtained with
an automatic interpretation system for AVHRR (Advanced Very High Resolution
Radiometer) infrared images, developed to detect and to label oceanic structures. This
automatic interpretation system contains mainly a set of different classifier, that allow
to compare the results. The classifiers use functional and operational principles de-
rived from the ANNs (Artificial Neural Nets), Rule-Based Knowledge Systems and
Fuzzy System.
Fig. 1 depicts the overall structure of the system that has been developed. In a first
step, the raw image is processed by means of algorithms such as radiometric correc-
tion, map projection and land masking. The second step aims to detect clouds pixels
that are opaque to radiance data measured in the AVHRR infrared and visible
scenes[1]. The following task is the segmentation that will divide the whole image in
regions. The nature of ocean dynamics makes very difficult this process that is never-
theless fundamental, so we’ve designed an iterative knowledge-driven method to
perform this part of the process pipeline [2]. The next task is the features or descrip-
tors selection, which consists of selecting an optimal or sub-optimal feature subset
from a set of candidate features. The most common framework for features selection
is to define some criteria for measuring the goodness of a set of features [3], and then
use a search algorithm to find an optimal or sub-optimal set of features [4]. We have
used Bayesian networks for features selection in the recognition of oceanic structures
in satellite images [5][6]. In the last step, each region produced in the segmentation is
analyzed and, if the recognition is positive, it is labeled with the identifier of the
matching structure. The structures of interest in the Canary Islands zone as defined in
[7] are: coastal upwelling, warm eddies, cold eddies and island wakes.
Fig. 1 Structure of the ocean feature recognition system.
We have implemented a redundant recognition subsystem. It has an ANN-based

Symbolic Processing Element (SPE) module, a rule-based Graphic E.S. (GES),
Bayesian Network, Hybrid System (Artificial Neural Network based Radial Base
Function and Fuzzy System based Sugeno [8]) and Neuro-Fuzzy Systems
(NEFPROX, ANFIS [9] [10]) performing the same task. The purpose is to test differ-
ent methodologies and to provide a way to validate and compare these results.
We have explored the use of Bayesian networks as a mechanism for feature selec-
tion in a system for automatic recognition of ocean satellite images. The use of
Bayesian networks has provided benefits with respect to SPE, not only in the reduc-
tion of relevant features, but also in discovering the structure of the knowledge, in
terms of the conditional independence relations among the variables. In future works
we plan to improve the accuracy rate of the system including more variables. Fur-
thermore, we expect to use models to avoid the discretisation of the continuous fea-
tures when learning Bayesian networks.
On the other hand, obtained results aren´t expected for classification process by
hybrid and neuro-fuzzy system. They don´t correspond with results of other cassifiers
(EPS, GES and Bayesian Network). The main problem is the number of generated
rules, which is excessive. In future works we plan to improve the accuracy rate of the
hybrid and neuro-fuzzy systems including more variables, tunning training parame-
ters, optimizing the obtained structure and automating the construction of neurofuzzy
system.
Finally, the structure of the automatic interpretation system that has been intro-
duced, it is a complex and independent structure for the following tasks:
1. Iterative segmentation-recognition cycle is made by means of graphic expert sys-
tem.
2. Feature selection and validation of knowledge is done by bayesian learning
3. Multiple classification allows to find different interpretations of the existing
knowledge to validate the knowledge of the classifiers.
References
1. Torres J.A., Guindos F., Peralta M. and Cantón M.: An Automatic Cloud-Masking System
Using Backpro Neural Nets for AVHRR Scenes, IEEE Transactions On Geoscience and
Remote Sensing, vol. 41, nº 4, (2003) 826–831.
2. Guindos F., Piedra J.A. and Cantón M.: Ocean features recognition in AVHRR images by
means of bayesian net and expert system, 3rd International Workshop on Pattern Recogni-
tion in Remote Sensing, Kingston University, United Kingdom. August (2004).
3. Yi Lu Murphey and Hong Guo: Automatic Feature Selection – a hybrid statistical approach.
IEEE, (2000) 382–385.
4. Langley P. and Sage S.: Induction of selective Bayesian classifiers. In Proceedings of the
Tenth Conference on Uncertainty in Artificial Intelligence, Seattle. Morgan-Kaufmann,
(1994) 399–406.
5. Bruzzone L.: Classification of remote-sensing images by using the Bayes rule for minimun
cost. IEEE Int. Geoscience and Remote Sensing Symposium, (1998) 1778–1780.
6. Yamagata Y. and Oguma H.: Bayesian feature selection for classifying multi-temporal SAR
and TM Data. IEEE, (1997) 978–980.
7. Arístegui J., Sangrá P., Hernández-León S., Cantón M., Hernández-Guerra A. and Kerling
J.L.: Island-induced eddies in the Canary Islands, Deep-Sea Research, nº 41, (1994) 1509–
1525.
8. Zadeh L.A.: Fuzzy Logic. IEEE Computer, vol. 21 , issue 4 , April (1988) 83–93.
9. Roger Jang: ANFIS Adaptative Network based Fuzzy Inference System. IEEE Transactions
On Systems, Man and Cybernetics, vol 23, nº 3. May/June (1993).
10. Nauck D. and Kruse R.: A neuro-fuzzy approach to obtain interpretable fuzzy systems for
function approximation. Fuzzy Systems Proceedings, IEEE World Congress on Computa-
tional Intelligence., vol 2 , May (1998) 1106 –1111.
From Projective Virtual Reality to the Virtual Human:
Virtual Worlds become realistic
Jürgen Rossmann
Institute of Robotics Research, Otto-Hahn-Str. 8, 44227 Dortmund

{rossmann@irf.de}
http://www.cosimir.com
Extended Abstract
Latest experiences in industrial and space applications have shown, that the symbiosis
of virtual reality techniques with versatile robot control techniques can break new
grounds in the field of virtual reality applications as well as in the field of robotics. In
this context the new paradigm of Projective Virtual Reality is presented as a
sophisticated new approach to the control and supervision of robotic and automation
systems.
Fig. 1. The basic idea of Projective Virtual Reality
The idea behind Projective Virtual Reality is to "project" actions that are carried out
by users in the virtual world into the real world, primarily by means of robots but also
by other means of automation. Robot control technology thus provides the user in the
virtual reality with a "prolonged arm" into the physical environment, paving the way
for a new quality of a user-friendly man machine interface for automation
applications. The paper provides the background of Projective Virtual Reality which
is based on latest results in the fields of task planning, world modeling, cooperative
robot control and sensor-based robot control. Building on these basic ideas, new
research directions related to simulating the “virtual human” will be presented. The
implemented approach of the real-time dynamic simulation of an anthropomorphic
kinematic chain will be demonstrated. The implemented virtual human in the
applications presented not only serves as an ergonomic “role model” in a digital
factory environment, but it is also flexible enough – and easy enough to use – to be
used for general scientific applications ranging from the control of anthropomorphic
agent, the display of emotions by body language and various medical applications.
Fig. 2. The dynamic simulation of an astronaut on board the International Space Station
The first step to make virtual worlds “feel” realistic was to simulate the underlying
physics. We are currently pursuing the next step which is to populate these “physical
virtual worlds” with human shaped avatars to further increase user-acceptance. Will
we be able to simulate a virtual society some time soon?
Fig. 3. Dynamic interaction between multiple avatars in a space laboratory environment

Multimedia Infrastructure for Intelligent Diagnosis and
Processing of Phonocoradiographic Signals?
J. Martı́nez1 , J. Palma2 , M. Breis2 , K. Herzog3 , A. Fritz3 , and J. Lopez4

1
Dpt. of Electronics, Computer Technology and Projects.
Technical University of Cartagena. Spain.
juanc.martinez@upct.es
2
Artificial Intelligence and Knowledge Engineering Group.
University of Murcia. Spain
jpalma@dif.um.es
3
Software Engineering for Medicine.
FH-Hagenberg. Austria
{katharina.herzog,andreas.fritz}@fh-hagenberg.at
4
Murcia General Hospital. Murcia. Spain
Extended Abstract
The traditional method of auscultating heart sounds is still the first basic analysis tool
that is applied in the evaluation of the functional state of the heart, as well as the first
indicator used to submit the patient to a cardiologist. A well-trained medical doctor can
reach a final diagnosis immediately after an exploration and a conscientious ausculta-
tion. During the last 15 years, heart auscultation has undergone the spectacular advance
of other high-technology diagnosis techniques such as the doppler-ultrasound transtho-
racic echocardiogram (TTE). Nevertheless, the needed high-technology devices can be
exclusively found at high-level hospitals, while it is absent in most cardiological ambu-
latory services as well at Primary Care Centers for adult and pediatric population.
The loss of importance of heart auscultation as a technique of exploration has con-

ducted at the same time to a backward movement in its teaching as family practice
trainees in both undergrade and postgraduate levels, and as a consequence it has leaded
to an important lack of auscultatory skills in the medical staff [1]. This lack has given
rise to an increasing number of erroneous diagnoses in primary attention which are un-
necessaryly derived to cardiologist specialists [2]. Other additional factors that explain
this decreasing interest are related to: a) the subjective character of the auscultation and
its strong dependence on auscultatory experience; b) the lack of objective documenta-
tion to contrast the auscultatory findings found in the patient; c) and the poor frequency
response of the traditional commercial stethoscopes. Phonocardiography (PCG) was in-
troduced to avoid the limitations of auscultation technique. A Phonocardiogram allows
documenting timing, relative intensity, frequency, quality, tone and location of the dif-
ferent components of the heart sound, in an objective and repetitive way.
?
This work was supported by the Spanish MCyT under project MEDICI (TIC2003-09400-C04-
01) and Murcia Regional Government under project PB/46/FS/02
Nevertheless, in recent years, the medical community has shown a renewed interest
in cardiological auscultation and PCG. This renaissance of PCG is motivated by the
confluence of a diversity of factors: a) the present situation of overutilization of high
technology devices (such as TTE), which are restricted to a few Cardiology units, b)
the demonstrated usefulness of cardiac auscultation in clinical prediagnosis, especially
in primary attention, since some pathologies can be detected by this process. Further-
more, many pathological conditions of the cardiovascular system cause murmurs and
alterations in the sounds of the heart, even before they were detected as symptoms by
means of other more advanced techniques [3].
In this work, a system for the storage, visualization, automatic detection and di-
agnosis of alterations of the heart function, based on the analysis of the PCG signal, is
proposed. Automatic detection and diagnosis requires a previous characterization of the
PCG morphological properties, in order to proceed with the identification. The process-
ing of the PCG signal is base on a hierarchical structure that provides a methodology
for the analysis and interpretation of the phonocardiographic signal, with a scheme sim-
ilar to that followed by the physician during auscultation. This structure is based in the
abstraction of signals and organized in different levels, each one with different blocks
aimed to specific tasks, which perform the complete processing from the acquisition of
the phonocardiographic record to the final diagnosis. Besides the abstraction levels [4],
in each level several specialists exist, which are processing blocks that perform analysis
of the associated signals or extract some kind of information. Each specialist is defined
with input data, grouping operations, and output data, which will be available to the rest
of specialists of the same level and higher. The lowest level of the hierarchy performs
the most intensive computations, whereas as the levels are higher, the information is be-
ing abstracted progressively. Examples of methodologies like the proposed in this paper
have been applied successfully to other biomedical environments, like the detection of
myocardial ischaemia [5]. The proposed hierarchy is organized in four levels, each one
associated to the main signals: envelopes, events, sounds, and diagnosis.
The proposed system pursues a double goal. On the one hand that of serving as ob-
jective element for prediagnosis of alterations of the heart function in primary attention.
On the other hand, the second goal is focused on the exploitation of training capabili-
ties of the system. To achieve these goals, a multimedia database is provided in which
the PCG signal and the processing results are stored together with the information of
patient’s clinical history. In this multimedia database, different metaphors for the PCG
visualization are offered, as well as the corresponding TTE video, thus making possible
for the user to associate PCG sounds with real heart images. Finally, an important effort
has been made in providing the medical stuff with an adequate system for the acqui-
sition and electronic storage of the PCG signal. In this sense, an ubicuous acquisition
system, based on PDA, has been developed. This PDA acquisition systems allows the
medical stuff to carry out the auscultation process, and at the same time, allows them to
add annotations to acquired the PCG signal.
References
1. Mangione, S., Nieman, L.: Cardiac auscultatory skills of internal medicine and family practice
traineers. Journal of the American Medicla Association 278 (1997) 717–722
2. Gaskin, P., Owens, S., Talner, N., Sanders, S., Li, J.: Clinical auscultation skills in pediatrics
residents. Pediatrics 105 (2000) 1184–1187
3. Lukkarinen, S., Noponen, A., Angerla, A., Sikiö, K., Sepponen, R.: A recording stethoscope
in study of heart sounds of children. Medical & Bilogical Engineering & Computing 34 (1996)
97–98
4. Milios, E., Nawab, S.: Signal abstractions in signal processing software. IEEE Trans. on
Acoustics, Speech, and Signal Processing 37 (1989) 913–928
5. Ruiz, R., Barro, S., Presedo, J., Palacios, F., Vila, J.: Abstraction of information in elec-
trocardiographic monitoring for the detection of myocardial ischaemia. In: Proc. of the V
International Symposium on Biomedical Engineering. (1994) 291–294
Dactylology Language Translator from Video to Text
Alcorac Camino, Carlos M. Travieso, Jesús B. Alonso and Miguel A. Ferrer
Dpto. de Señales y Comunicaciones, Universidad de Las Palmas de Gran Canaria

Campus de Tafira s/n, E-35017, Las Palmas de Gran Canaria, Spain.
e-mail:ctravieso@dsc.ulpgc.es
Abstract. In this paper an innovative system has been created, which detects
the movement of the hand and implements a converter of dactylology language
to text, segmenting an input video and recognizing the frames selected, by
means of an algorithm of a movement detector system. Once, it established the
correct images, these have been parameterizated and recognized with a decision
tree classifier and a network neuronal (NN), from the characteristics of the
different classes of the dactylology alphabet, with a success of the 93.9%. The
reached times in these tests show their possible application in real time.
1 Introduction
Since the principle of the times, the human species has gone feeling the need to have a
greater quality of life. Along the history, the technological development has achieved
an evolution that contributes clearly to create a welfare society as we find it
nowadays. In spite of the conscience of global social integration, it exists minorities
represented by collective or groups that cannot opt to an improvement of the quality
of life, since they do not possess the same options neither opportunities that the
remainder of the citizens. The present associate-cultural tendency produces an
accelerating continuous in the need of routine interaction of the users with diverse
systems, which take advantage to try to break the barriers that impede the integration
of these groups [1]. For this reason, the motivation of this work comes given by the
application of a system to recognize the movement of the hand (video) to text, with
the purpose to be able to break the barrier of language among the people with auditory
decrease and the remainder of the society.
Captured video
Web cam
Fig. 1. Block diagram for the proposed system

Subsequently, this process to carry out the dactylology language converter system
will be described. This implemented system recognizes the symbols that compile in
the Spanish dactylology alphabet, which is utilized in the language of signs for the
communication of the disabled, considering the dynamic conditions of the
gesticulation of the above language to select adequately, the gestures represented in
the gesticulation, and to carry out the conversion of these gestures to text (see figure 1).
2 Movement detection
To carry out this conversion we need a device sensor (web camera, video camera
more acquiring card, digital or even, a mobile telephone with camera integrated) for
receiving the images that contain the gesticulation.
To obtain a capture of the correct gesture has been utilized a uniform background,
for the acquisition of the video samples, complying a series of environmental
conditions as the brightness, the position of the hand on the background, etc. Also,
some own parameters of the capture are established among them, we emphasize the
shutter speed of the input device, which has been acquired to 30 frames by second.
The extraction of the frame (of the input video sequence), pertaining to a definite
gesture, is carried out by the monitoring of the movement of the hand, utilizing a
movement detector system (SDM), designed to consider the dynamic conditions of
the gesticulation.
In the above algorithm (SDM), as the first step is carried out a test of the video
each 4 frames, this signifies that the sequence is analyzed each 0.13 seconds. On this
way, we will be grasped with security all the possible gestures carried out. From each
frame extracted, some peripheral parameters are calculated that give an idea of the
position of the hand in that determined instant. These data are stored to do a
comparison with the same data but of the following frame. Thus, we can have an idea
of the movement carried out by the hand, in the transition between the two captures
(elapsed 0.13 seconds). For it, two types of movements are defined, depending on the
comparison of the peripheral distances calculated with respect to some thresholds
created for the operation of the SDM:
– Abrupt Movement: is produced when the hand carries out a considerable

displacement in an interval of 0.13 seconds.
– Light Movement: when the displacement of the hand corresponds with a small
change of the position of the fingers, a small rotation of the hand, etc.
When a static gesture is detected, with respect to the previous frame, is selected in
order to extract its characteristics for its subsequent classification. The remainder of
the frames are ruled out as transition or stable frames due the static position of the
hand with the time. It is possible to capture the same gesture more than one time,
since the hand can fluctuate of natural form or, even, the gesticulation can carry out
repercussions due to the inexperience of the use of the language of signs on the part of
the person that makes the shot. For this reason, the segmented gestures have been
studied to avoid a repetition of the same, ruling out those whose characteristics are
very similar. At the same time, it will serve to do a classification that summons to the
gestures to some specific groups depending on its morphology.
Variability
Instable Cycle Stable Cycle e

Abrupt
Movement
Light
Movement
Stable
0.13 0.26 0.39 0.52 0.65 Seconds
Ruled out Ruled out Selected Ruled out Ruled out

Fig. 2. Detection movement process
Achieving it, some general characteristics are extracted, which gives an idea of the
dimensions, area and solidity of the gesture. Subsequently, perimeter recognition is
done to find more specific parameters, as the number of fingers, inclination of the
hand, etc. Attending to these characteristics calculated will proceed to a classification
in which the gestures with similar properties establish groups that define the type of
gesture. The creation of the groups is carried out to obtain an organization of the
alphabet in the way of tree; this causes a faster recognition of the segmented frames.
Analyzing the alphabet are found gestures, whose similarity is evident, as it is the
case of the letters ‘s’ and the ‘o’. Solving it, some more specific parameters are
extracted, which help to recognize the relation among the similar gestures.
3 Experiments: Parameterization and Classification
To extract the characteristics of the gesture processed [2][3] and to assign them for
a classification, a decision tree classifier is created gestures for the all the gestures of
the alphabet. In some cases the sub-classification of the alphabet can go to the
creation of the exclusive groups for a single gesture, in such case, a direct answer is
obtained, which do not need that the gesture be processed to take it like input of the
recogniser.
The characteristic parameters of each gesture segmented are stored in sequential
form, to carry out an exhaustive comparison of these parameters, so much the generic
as the specific parameters. On this way, we can know that gestures must be
discriminated and which ones must be recognized. Therefore, this system is able of
detecting the characteristics of two similar consecutive gestures, in this case, the last
processed frame will be ruled out. If they are different, the two frames will be
analyzed since they are taken like valid and different samples.
The gestures of the input video, that are not recognized directly, need a pre-
processing in order to prepare the data to the input of the neuronal networks (NN) [4],
which form the classifier based on the groups of letters, that are established in the
characteristics extraction.
For the analysis and evaluation of this system are needed two databases. A
dynamic database, that is included for short videos where are represented the
variability of the gestures of the hand, and the form of the movement for its
segmentation is verified. And another one is static [5], which serves for training and
test the NN [4]. In the following figure, we can see one sample of each gesture.
Fig. 3. Samples from our database
4 Conclusions
To conclude, in this paper have created an innovative system that detects the
movement of the hand for the Spanish dactylology language, that is capable of
segmenting an input video and to recognize the frames selected by a classifier based
on decision tree of the dactylology alphabet and a NN [4]. This system reaches an
average success rate of 93.9%, with a low computation cost in all their steps, which is
possible suitable for real time applications (compiling the programs implemented
from MATLAB to C language), therefore, this system permits to do an online
translation of the gesticulation.
5. Acknowledgement
This work has been developed and agrees by research Project UNI2003/26 from
University of Las Palmas de Gran Canaria (Spain).
References
1. Gascón-Ricao A.: http://www.ucm.es/info/civil/bardecom/docs/signos.pdf. ¿Señas o
signos?: Evolución histórica. [Active on October 2004]
2. O’Gorman L., Kasturi R.: Document image analysis. IEEE Computer Society Press, 1996
3. “Image acquisition toolbox”, User’s Guide, Version 5, The Math Works, INC, (2004)
4. Bishop C.M.: Neuronal network for pattern recognition. Ed.Oxford University Press,
(1994)
5. Travieso C.M., Alonso J.B., Ferrer M.A.: Sign Language to text by SVM. Proceedings of
Seventh International Symposium on Signal Processing and its Applications, vol. 2,
(2003) 439-442.
Automatic Braille Character Recognition: BrailLector
Néstor Falcón, Carlos M. Travieso, Jesús B. Alonso and Miguel A. Ferrer
Dpto. de Señales y Comunicaciones, Universidad de Las Palmas de Gran Canaria

Campus de Tafira s/n, E-35017, Las Palmas de Gran Canaria, Spain.
e-mail: nestor@cibeles.teleco.ulpgc.es; ctravieso@dsc.ulpgc.es
Abstract. In this paper we present the development of BrailLector, a system

able to speak from Braille writing. By means of dynamic thresholding, adaptive
Braille grid, recovery dots techniques and TTS software (Text-To-Speech),
BrailLector translates Braille scanned images into normal text, and not only
that, it speaks the translated text. BrailLector is a robust application with
innovative thersholding and Braille grid creation algorithms which detects and
read Braille characters with 99.9% of correct symbols and an error variance
below 0.012. The conversion time is only 26 secs for double-sided documents
by MATLAB programming language.
1. Introduction
Nowadays, lack or problem of vision has been an important obstacle to access to
printed contents and to the information society. For this reason, some people have
tried to achieve that blind people are able to access to the printed culture, for example
Valentin Haüy and Luis Braille [1] who have understood the importance of a
communication code. Globally, an estimated 40 to 45 million people are blind and
135 million have low vision according to the World Health Organization (WHO) [2]
and this number grows every year.
A Braille Optical Character Recognizer is interesting due to the following reasons;
• It is an excellent communication tool for sighted people (who do not know

Braille) with the blind writing.
• It is a cheap alternative Braille to Braille copy machine instead of the current
complex devices which use a combination of heat and vacuum to form Braille
impressions.
• Braille writing is read using the finger so is necessary touch the document, for
this reason the book after many readings is possible has been deteriorated.
• It is interesting to store a lot of document of blind authors which were written in
Braille and were never converted to digital information.
Since the most part of Braille books are written with two kinds of points, we will
have to distinguish between each one. One kind is like “mountains” and the other is as
little “valleys”, hence the finger of the blind person who read the text only detects the
protrusions, while depressions are points written to be read from the other side of the
sheet. If we are able to distinguish between each point we will have a big advantage,
we will be able to recognize the two sides with only one scan.
Fig. 1. Protrusions and Depressions on a Braille sheet
The structure of this paper is the next: In the next text we will describe the
characteristics of the database created. After it will be described image processing
techniques used for dots detection and recovering and Conclusions closes this
extended abstract.
2. Database and Image Processing

A big database has been created in order to check the global system with as many
characters as we could. This data base provides single and double-sided documents
and its size ensures the correct testing of the developed system and a good analysis of
the error variance. The next table gives a full explanation of this database.
Table 1. Database created
Braille sheets 26
Total number of characters 30862
Mean number of character per sheet 1235
Digital format Gray scale
Resolution 100 dpi (horizontal and vertical)
Image size 1360 Kbytes
Image format Bitmap (‘bmp’)
Braille type Double sided – grade 1
Document size 29.5 cm. (horizontal) x 30.5 cm. (vertical)
The mechanism used for image acquisition has been a flat-bed scanner because it is
a cheap alternative which can be used for so many other applications and it is easy
and quick to use.
The different steps of the scanned image for its translation are;
1. An innovative thresholding method has been used to extract the useful

information of Braille images. An iterative algorithm looks for the best threshold
according to the areas of Braille dots. This area criterion for thresholds selection
offers an accuracy way to get the optimum levels to separate black, white and
grey.
Once this primary process is done, we take advantage of the shadows which
make dot patterns. As we have seen that protrusions on a Braille sheet have a
brilliant zone above and a dark zone below. The depressions have exactly the
opposite pattern (these shadows are created by the skew angle of the light beam
in reflection scanners) so this difference will let us to separate front and back side
dots;
o Moving white “islands” 4 pixels downwards and doing a logical “and” with
black spots we will extract front side dots.
o Moving white “islands” 4 pixels upwards and doing a logical “and” with
black spots we will extract back side dots.
2. The goal of this secondary process is to separate each side of the document in
different images. This algorithm is the fastest we have tried since it avoids the
sequential reading of the image matrix and it is very simple and efficient. In this
stage, skew angle of the scanned document is detected by means of horizontal
histogram and mass centers calculation and corrected rotating the original image.
3. Not all the dots are detected with overlapping process, some of them are missed,
either because they are very small or their shadows do not overlap with only 4
pixels. For this reason, a new stage was added to the system: The Braille grid.
An adaptive algorithm has been developed in order to make this mesh from the
detected dots. Since distances between points are normalized [3], the algorithm
builds columns and rows following these distances and according to the detected
dots until now. This grid is flexible and respects the layout of the original
document which makes it suitable for copying Braille sheets without losing the
format.
Fig. 2. Braille grid layout for frontal side
4. After mesh building, all valid Braille positions are known. Now we fit this image
on the dots image after thresholding in order to check in detail those positions
where dots were not found. In this point lays the intelligence of the global
system, only potential positions of dots are checked; Original dots will be
recovered (they belong to correct Braille positions) and false dots will be
discriminated as they are out of the valid places for a Braille dot.
The final image with the detected dots is analyzed and text is segmented in rows
and characters. Every character will be converted into a binary number according
to the active dots.
Fig. 3. Standard Braille Character (left) and ‘r’ symbol

In this way and looking the previous example, ‘r’ symbol can be coded like
‘111010’ where each black dot is an active dot or a “mountain” on the paper and they
become ‘1’ in the binary number. This way of Braille text binarization makes the
global system independent of the language of the document and easily configurable
for adding different alphabets.
3. Conclusions
Finally, in this paper we have developed an automatic system for translating Braille
text to normal text or voice. The global algorithm is very fast and robust. For
achieving this system, dynamic thresholding and adaptive Braille grid has been used,
adding some intelligence to the global process and making it able to detect dots in
both sides of the document with only one scan. This process has an efficiency of
99.9% and it takes only 26 secs to translate a double-sided document improving all
the main references found in the bibliography [4],[5],[6].
Fig. 4. Braille Image Processing
4. Acknowledgement
This work has been developed and agrees by research Project UNI2003/26 from
University of Las Palmas de Gran Canaria (Spain).
References
1. http://www.asac.org.ar/Publi/bra_b001.htm Active on October 1st 2004
2. http://www.who.int Active on October 1st 2004
3. Dubus J., Benjelloun M., Devlaminck V., Wauquier F., Altmayer P.: Image processing
techniques to perform an autonomous system to translate relief Braille into black-ink,
called: Lectobraille. In Proceedings of the Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, (1988) 1584–1585.
4. Wong L., Abdulla W., Hussman S.: A Software Algorithm Prototype for Optical
Recognition of Embossed Braille. The University of Auckland, New Zealand, (Technical
Report) (2004).
5. Mennens J.,. Tichelen L. V, Francois G., Engelen J.: Optical recognition of braille writing
using standard equipment. IEEE Transactions on Rehabilitation Engineering, 2(4) (1994)
207– 212.
6. Ng C., Ng V., Lau Y.: Regular feature extraction for recognition of Braille. In Proceedings
of Third International Conference on Computational Intelligence and Multimedia
Applications, (1999) 302-306.
Retinal Based Authentication via Distributed Web
Application
C.Mariño, M.G.Penedo, and M.Penas
Grupo de Visión Artificial y Reconocimiento de Patrones (VARPA)

Universidade de A Coruña,
Campus de Elviña s/n,
castormp@fi.udc.es,{cipenedo,infmpc00}@dc.fi.udc.es
Abstract. Traditional authentication systems, employed to gain access to a pri-

vate area in a building or to data stored in a computer, are based on something the
user has (an authentication card, a magnetic key) or something the user knows
(a password, an identification code). But emerging technologies allow for more
reliable and comfortable for the user, authentication methods, most of them based
in biometric parameters. We have developed a new methodology for personal au-
thentication, where the biometric parameter employed for the authentication is
the retinal vessel tree, acquired through a retinal angiography. In this work we
will present the design and implementation stages of an application which allows
for a reliable personal authentication in high security environments based on the
retinal authentication method.
1 Introduction
Reliable authentication of people has long been an interesting goal, becoming more im-
portant as the need of security grows, so that access to a reliable personal identification
infrastructure is an essential tool in many situations (airport security controls, all kinds
of password-based access controls, ...). Conventional methods of identification based
on possession of ID cards or exclusive knowledge are not altogether reliable. ID cards
can be lost, forged or misplaced; passwords can be forgotten or compromised. A solu-
tion to that problems has been found in the biometric based authentication technologies.
A biometric system is a pattern recognition system that establishes the authenticity of
a specific physiological or behavioral characteristic possessed by a user. Identification
can be in the form of verification, authenticating a claimed identity, or recognition, de-
termining the identity of a person from a database of known persons (determining who
a person is without knowledge of his/her name). Many authentication technologies can
be found in the literature, with some of them already implemented in commercial au-
thentication packages [1–3]. But today most of the efforts in authentication systems
tend to develop more secure environments, where it is harder, or ideally, impossible,
to create a copy of the properties used by the system discriminate between authorized
individuals and unauthorized ones, so that an impostor could be accepted by the bio-
metric system as a true sample. In that sense, the method employed by our system [4]
uses for authentication biometric parameter the blood vessel pattern in the retina of the
eye.
In this paper, a distributed client-server framework which allows for personal au-
thentication is outlined. From the design stage, where design patterns are the funda-
mental tool, to the implementation (server,communication channels and clients) are de-
scribed. The main goal of this work is to build a robust system which allow for the
automatic personal authentication, which would allow or deny the access to critical
resources of the system.
2 Description of the distributed system

Our system employs a conventional client-server architecture [5]. Once the retinal pat-
tern of the authorized person is stored in the database, the system performs an authen-
tication procedure each time the individual try to access the protected resources of the
system. Figure 1 depicts a schema of the authentication process.
Web/application server
Authentication server
Authentication HTTPS XML/RPC

request
HTTPS
Acceptance/
rejection
Users database
Fig. 1. The distributed authentication system.
The more important tools employed for designing the system has been the design
patterns [6] and the UML [7, 8], which allow for a more robust and reliable software
applications.
A client application initiates a connection to authenticate an individual, and if au-
thentication is successful, resources can be accessed. The graphical user interface for
the interaction with data (like the depicted in Figure 2) is obtained through the web
application.
The overall operation flow at the client-side can be described as follow (see also
Figure 1):
– An individual makes HTTP request to web server first. The web server will handle
the request and will communicate with the authentication server to perform the
authentication process.
– The authentication server will access the database and will allow or deny the access
to the requested resource. Upon request, procedure execution will be invoked (using
XML-RPC) based on the data sent by the client.
– Once the procedure ends, the related information is sent to the client and displayed
in the browser window, allowing or denying access depending on the result of the
authentication process.
Fig. 2. Screen from the client program-application with retinographies, which are the biometric
authentication parameter (a) and obtained authentication parameters (b).
3 Communication technology
To assure confidentially and reliability in the communications, the most widely used
secure Web encryption has been employed: the Secure HyperText Transport Protocol
(HTTPS) on the Secure Socket Layer (SSL) version 3.0 with 40 and 128 bit RC4 en-
cryption. The SSL [10] protocol has become most widely used method for encrypting
and authenticating Web communications.
Our approach to deal with the heterogeneity of the clients focuses on documents or
data and transforms that visualizes the data to fit the computing capabilities of differ-
ent users, while preserving semantics of data. This division of the underlying data and
the way it gets displayed mimics the Model-View-Controler (MVC), a well known and
frequently used design pattern to develop interactive applications with flexible human-
computer interfaces. This would allow for a future implementation of mobile authenti-
cation clients, for example.
XML has been chosen as the medium to represent the data sent to the server for
the authentication. XML is a markup language for documents containing structured
information. Once the data is written in XML, as associated XSL document can be
written to define the way it gets displayed. The XML document is platform-independent
and thus corresponds to the model, whereas the style sheet depends on the displaying
device, and corresponds to the view.
MVC separation of view and model offers the following advantages over conven-
tional application-centric environments:
– Model-view separation at the document level via XML/XSL documents.

– Simple communication protocol with a standard message format based on (en-
crypted) ASCII XML messages.
Finally, using XML-RPC [11], procedure calls are wrapped in XML establishing
that way a simple pathway for calling authentication functions. XML-RPC is a quick-
and-easy way to make procedure calls over the Internet. It converts the procedure call
into XML document, sends it to a remote server using HTTP, and gets back the re-
sponse as XML. Other considered (although finally discarded) protocols where SOAP
and CORBA, both of them popular protocols for writing distributed, object-oriented ap-
plications, well-supported by many vendors, although they are much more complex than
XML-RPC,and require fairly sophisticated clients. In addition, XML-RPC has fewer
interoperability problems than them.
4 Conclusions and future work

In this paper an authentication web application has been presented. Using that program
restricted access to system resources (buildings, data, etc.) can be granted only to au-
thorized individuals, rejecting those requests from unauthorized persons. XML, which
serves as the communications medium, is a standard that has already gained wide accep-
tance and provides a powerful medium for data exchange, visualization specifications
and procedure execution by means of the XML-RPC method invocation procedure.
Moreover all the communications take place over secure channels, through to the em-
ployment of encryption with HTTPS on the SSL version 3.0. Currently, system is under
evaluation, and many tasks as robustness assessment, performance verification, formal
verification of whole the server-procedures, so as testing in a real environment must be
performed.
Acknowledgments
This paper has been partly funded by the Xunta de Galicia through the grant contract
PGIDIT03TIC10503PR
References
1. J.G. Daugman. Biometric personal identification system based on iris analysis. United States
Patent No.5,291.560, 1994.
2. J.Bigüin, C.Chollet, and G.Borgefors, editors. Proceedings of the 1st.International
Conference on Audio- and Video-Based Biometric Person Authentication, Crans-
Montana,Switzerland, March 1997.
3. R. Zunkel. Hand geometry based verification. In BIOMETRICS:Personal Identification in
Networked Society. Kluwert Academic Publishers, 1999.
4. C. Mariño, M.G. Penedo, and F. González. Retinal angiographies based authentication. Lec-
ture Notes in Computer Science, 2905:306–313, 2003.
5. Robert Orfali, Dan Harkey, and Jeri Edwards. Client/Server Survival Guide. John Wiley &
sons, 3rd edition, 1999.
6. E. Gamma, R. Helm, R. Johnson, and Vlissides J. Design Patterns, Elements of Reusable
Object-Oriented Software. Professional Computing Series. Addison-Wesley, 1995.
7. Mark Grand. Patterns in Java: a catalog of reusable design patterns illustrated with UML,
volume 1. New York,John Wiley & sons, 1998-1999.
8. Mark Grand. Patterns in Java: a catalog of reusable design patterns illustrated with UML,
volume 1. New York,John Wiley & sons, 1998-1999.
9. Java 2 sdk, standard edition version 1.3.1, plug-in installation notes.
http://java.sun.com/j2se/1.3/install-linux-sdk.html.
10. K.E.B. Hickman. The SSL protocol. http://www.netscape.com/newsref/ssl.html, December
1995.
11. Simon St.Laurent, Joe Johnston, and Edd Dumbill. Programming Web Services with XML-
RPC. O’Reilly, 2001.
Medial Axis Transform Computation Using
Shock Wavefront Detection
Rubén Cárdenes1 and Juan Ruiz-Alzola1,2

1
Medical Technology Center, University of Las Palmas GC, Spain
2
Canary Islands Institute of Technology (ITC), Spain
1 Introduction
The medial axis transform (MAT) or skeleton of a geometric figure is a set of
curves that approximate the local symmetry axis of this figure. The concept of
medial axis consists in the reduction of geometric shape information to the min-
imum. The first definition was given by Blum [1], who stated the analogy with a
prairie fire: the MAT is the set of points where the propagation wavefront initi-
ated from the shape boundary “intersect itself”. Despite the extreme simplicity
of this definition, the implementation in digital images without loss of important
properties is surprisingly difficult. For this reason there exists a large number of
algorithms and methods in the literature to address this problem. The relevance
of MATs was introduced initially by Blum mainly for visual perception analysis
and it is used especially for pattern recognition.
Formally there exist four definitions for medial axis transform, all of them
equivalent, described by Montanari in [2]. In the first of them the medial axis
transform is the prairie fire model proposed by Blum commented above where
the fire lines represent the propagation fronts. The second definition states that
the MAT is the set of ridges of the distance map constructed from the shape
boundary. In the third definition, one of the more common, the MAT is the set of
centers of the maximal discs of the figure, where a maximal disc is a cicumference
contained in the shape for which there exists no other circumference inside the
shape that contains it. The last model define the MAT as the set of points
that do not belong to any straight line segment connecting other points to their
respective closest boundary points.
The principal MAT properties are the connectivity, the preservation of the
main topological features of the shape, i.e. to not lose any main branch in the
skeleton and to not add irrelevant branches to it, and the Euclidean equidistance
of every point of the skeleton to two or more points on the boundary. Regarding
the discrete MAT, it is desirable to obtain one pixel thickness results to accom-
plish the narrowness property. Obviously in a discrete domain the equidistance
property is not possible to obtain, so it is necessary to relax this property to get
the most approximated result.
2 Method
Methods to construct MAT are well described in the literature [3–5]. In this
paper we propose a method to extract the MAT from a 2D shape, where the
distance map from the shape boundaries is computed using a propagation scheme
and then, the MAT is constructed detecting the points where the collision from
the wavefronts occurs.
The way of detecting the points where the wavefront intersection takes place
is carried out by a non propagation criterion. The discrete propagation fronts
initiated from the shape boundary, are implemented with a list of pixels at Eu-
clidean distance d from it, where d is an integer number, and these pixels are the
ones which have real distance values closer to d inside the shape. The wavefront
at distance d from the boundary will generate the wavefront at distance d + 1,
as long as there exist more inner non visited pixels. When a pixel can not be
propagated deeper into the shape, it is automatically labeled as member of the
skeleton. With this criterion the MAT obtained is not connected and has not
one pixel thickness.
The connectivity is not assured because two effects. The first one appears
because points which should be labeled as belonging to the MAT, are not de-
tected in cases where the propagation fronts coming from opposite directions do
not intersect at the same pixel but adjacent ones, as in figure 1(a). The second
effect happens when the shape becomes wider as illustrated in figure 1(b). The
first effect is corrected a posteriori using the distance map computed, connecting
the pixels of the computed MAT through the points with maxima values in the
distance map. The second effect is solved in the propagation process, by con-
necting the extremal points of the MAT computed in d (if any) with its nearest
point in the front d + 1. The narrowness property is easily achieved by a regular
d=3
d=2
d=2
d=1 d=1
d=0 d=0
(a) (b)
Fig. 1. Disconnection effects: wavefronts coming from opposite directions do not inter-
sect at the same pixel but adjacent ones (a), and aperture zones (b)
thinning which is highly efficient because only a small percentage of the image
has to be scanned, and only one thinning iteration is needed.
3 Results and Conclusions

We present some results in figure 2, where MAT from different shapes are suc-
cessfully obtained with our method. This algorithm presents several advantages.
First, the MAT is computed very efficiently because it uses an ordered prop-
agation scheme [6] which is a very fast operation. The post-processing step,
connection and thinning, is computed also very fast because a very small num-
ber of points are involved and only one thinning iteration is needed. Second,
the topological properties of the shape are always preserved as shown in our
results. Third, irrelevant branches in the MAT do not appear, i.e. it is not very
sensitive to boundary noise, and no pruning of the resulting skeleton is needed.
And finally this method can be also extended straightforwardly to 3D shapes,
obtaining the medial surface transform.
Fig. 2. Medial axis transform in several shapes
Acknowledgments
This work has been partially funded by an European grant FP6-507609 from
the European Commission, and has been partially supported by the Spanish
Ministry of Science and Technology and European Commission, co-funded grant
TIC-2001-38008-C02-01.
References
1. H. Blum, “A transformation for extracting new descriptors of shape,” in Proc.
Models for the Perception of Speech and Visual Form, Weiant Wathen-Dunn, Ed.,
Cambridge, MA, November 1967, pp. 362–380, MIT Press.
2. U. Montanari, “A method for obtaining skeletons using a quasi-euclidean distance,”
Journal of the Association for Computing Machinery, vol. 15, no. 4, pp. 600–624,
1968.
3. C.J. Ammann and A.G. Sartori-Angus, “Fast thinning algorithm for binary images,”
Image Vision Comput., vol. 3, no. 2, pp. 71–79, 1985.
4. R. Ogniewicz and M. Ilg, “Voronoi skeletons: Theory and applications,” in Proc.
IEEE Conf. Computer Vision and Pattern Recognition, CVPR, Los Alamitos, Cal-
ifornia, 15–18 1992, pp. 63–69, IEEE Computer Society.
5. R. Kimmel, D. Shaked, N. Kiryati, and A. M. Bruckstein, “Skeletonization via
distance maps and level sets,” Computer Vision and Image Understanding: CVIU,
vol. 62, no. 3, pp. 382–391, 1995.
6. B.H. Verwer, P.W. Verbeek, and S.T. Dekker, “An efficient uniform cost algorithm
applied to distance transforms,” IEEE Transactions on Pattern Analysis an Ma-
chine Intelligence, vol. 11(4), pp. 425–429, 1989.
Comparative Analysis of Calibration Methods
for a Static Camera
J. Isern González, J.D. Hernández Sosa, A.C. Domı́nguez Brito and A.M.
Naranjo Cabrera
Instituto Universitario de Sistemas Inteligentes y Aplicaciones Numéricas en

Ingenierı́a (IUSIANI)
Spain, LAS PALMAS 35017
{jisern, dhernandez, adominguez}@iusiani.ulpgc.es, angel@jqbit.com
Extended Abstract
In computer vision, camera calibration is a procedure that tries to know how a

camera projects a 3D object on the screen. This process is necessary in those
applications where metric information of the environment must be derived from
images. Many methods have been developed in the last few years to calibrate
cameras, but very few works (i.e. Tsai [10], Salvi and Armangué [8], Lai [7] or
Isern [5]) have been done to compare such methods or to provide the user with
hints on the suitability of certain algorithms under particular circumstances.
This work presents a comparative analysis of eight calibration methods for
static cameras using a pattern as reference: Faugeras [4], Tsai [9] (classic and
optimized version), Lineal, Ahmed [1] and Heikkilä [6] methods, which use a
single view of a non-planar pattern; Batista’s method [3] which uses a single
view of a planar pattern; and Zhang’s method [11] which uses multiple views of
a planar pattern.
The results provided by every method have been compared and the accuracy
in the pattern points reconstruction and the stability of the model parameters
when the pattern is relocated or the camera setup varies have been analyzed. We
define the stability of a calibration as the constancy of the process results when
one parameter changes. Two types of stabilities can be defined: on the one hand,
if only extrinsic parameter are varied (pattern location) the values obtained from
calibration in the intrinsic parameters (internal setup of the camera) should be
constant; on the other hand, if only intrinsic parameters are varied the values
obtained from calibration in the extrinsic parameters should be constant.
This study has been carried out with real and simulated images and using,
when ever possible, the Matlab code made available by the authors of the meth-
ods on the WWW. Different kinds of patterns (planar and non-planar) have been
used according to the exigences of each method. Also, the experiments have been
repeated using patterns with a different setup; i.e. the number of points and their
location have varied. The comparison of the results obtained by every method
has been carried out using the same reference images.
Working with real data, only the stability or global errors (2D-Error or 3D-
Error) can be studied for each method. We have labelled the mean distance
between real and estimated image coordinates and the mean distance between
real and reconstructed 3D coordinates of the pattern points, respectively, as the
2D-Error and the 3D-Error.
When the calibration is performed with real data, ground truth values of each
parameter of the camera model are unknown so it is impossible to determine that
accuracy with which these parameters are obtained. To overcome this difficulty
a simulator was developed.
Sensibility to different kind of noises sources has been studied to each method
and their results are compared due to the use of the simulator has allowed the
analysis of the influence of the determined noise separately of the rest.
The study of the stability of the calibration methods have been realized
both in the simulator and in real conditions. In order to analyze the stability of
extrinsic parameters a sequence of image was acquired making zoom, therefore
the focal length was varied, while the pattern remained static front to the camera.
This experiment has been repeated varying the location and setup (number of
points) of the pattern.
In order to analyze the stability of intrinsic parameters two types of ex-
periments have been realized. On the first one, the pattern has been displaced
parallel to the camera. On the second one, the pattern has been relocated in dif-
ferent positions and distances respect to the camera. In both cases the camera
setup remained constant.
The experiments have shown that most of these methods are not stable in
the sense that the set of intrinsic parameters returned by a calibration method
suffered important variations under small displacements of the camera relative
to the calibration pattern. Similar results have been obtained for extrinsic pa-
rameters when the camera only changed its internal configuration (i.e. when it
zooms in or out) and not its relative position to the calibration pattern. In both
cases the instability affects principally to the parameters that are directly related
with the element varied in the experiment. Therefore, when focal length varies
the pattern distance is very instable and viceversa. Similar results are obtained
for optic center when the pattern is displaced parallel to the camera.
On the other hand, the influence of the different types of noise has been study,
in concrete, noise in image 2D-coordinates, pattern points 3D-coordinates and
distortion of the image. This question has been analyzed only in the simulator
due to in real conditions the noises can not be aisled.
The error function minimized by the methods, usually the image disparity
(2D-Error), achieves similar error levels with all them, when the experiments
have been carried out in the simulator. This happens when the camera setup is
near to the ideal, that is to say, when the camera setup is near to guess values
introduced to some of the methods, still noise conditions are presented. However,
this has not resulted in similar parameter estimations among the methods and,
obviously, it has not guaranteed that the parameter estimations converge to the
ground truth values.
In general, the experiments have shown that the values of the parameters
obtained in the calibration process depend on the method used. In other words,
it exists high differences among the estimated value for a determined parameter
obtained by each method, still the same image is used as reference to realize the
calibration process.
Finally, we can conclude that the review and analysis of the calibration meth-
ods considered in this paper have shown that there exists a strong coupling be-
tween intrinsic and extrinsic parameters. The main practical implication of this
fact is that, when a camera is calibrated with any of these methods, we are ”cal-
ibrating” the camera with just that pose, as the obtained intrinsic parameters
will change with pattern location or camera pose. Hence, the current practice of
calibrating a camera that will be moved around, to obtain just its set of intrinsic
parameters, must be reconsidered when an accurate calibration is paramount.
References
1. M. Ahmed, E. Hemayed, and A. Farag. Neurocalibration: a neural network that can
tell camera calibration parameters. In IEEE International Conference on Computer
Vision, pages 463–468, September 1999.
2. X. Armangu, J. Salvi, and J. Batlle. A comparative review of camera calibrating
methods with accuracy evaluation. In Proceedings of 5th Ibero-American Sympo-
sium on Pattern Recognition, SIARP 2000, pages 183–194, September 2000.
3. J. Batista, H. Araújo, and A.T. de Almeida. Iterative multi-step explicit camera
calibration. In Proc. Sixth International Conference on Computer Vision ICCV98,
volume 15, pages 709–714, January 1998.
4. Olivier Faugeras. Three-dimensional computer vision: a geometric viewpoint. MIT
Press, 1993.
5. J. Isern Gonzlez. Estudio comparativo de mtodos de calibracin y autocalibracin
de cmaras. PhD thesis, University of Las Palmas de Gran Canaria, Las Palmas,
Spain, July 2003.
6. J. Heikkila and O. Silvén. A four-step camera calibration procedure with implicit
image correction. In Proc. IEEE Conference on Computer Vision and Pattern
Recognition, volume 1, pages 1106–1112, June 1997.
7. Jim Z.C. Lai. On the sensitivity of camera calibration. Image and Vision Com-
puting, 11(10):656–664, December 1993.
8. J. Salvi, X. Armangué, and J. Batlle. A comparative review of camera calibrating
methods with accuracy evaluation. Pattern Recognition, pages 1617–1635, June
2002.
9. Roger Y. Tsai. A versatile camera calibration technique for high-accuracy 3d
machin vision metrology using off-the shelf tv cameras and lenses. IEEE Journal
of Robotics and Automation, RA-3(4):323–344, August 1987.
10. Roger Y. Tsai. Synopsis of recent progress on camera calibration for 3d machine
vision. The Robotics Review, pages 147–159, 1989.
11. Zhengyou Zhang. A flexible new technique for camera calibration. IEEE Transac-
tios on Pattern Analysis and Machine Intelligence, 22(11):1330–1334, November
2000.
Cue Combination for Robust Real-Time
Multiple Face Detection at Different
Resolutions?
M. Castrillón-Santana, O. Déniz-Suárez, C. Guerra-Artal, and J.

Isern-González
IUSIANI Universidad de Las Palmas de Gran Canaria

Spain
{mcastrillon, cguerra, odeniz, jisern}@iusiani.ulpgc.es
1 Introduction
The face detection problem, defined as: to determine any face -if any- in the
image returning the location and extent of each [7], seems to be solved, accord-
ing to some recent works [4] [6]. Particularly for video stream processing, these
approaches focus the problem in a monolithic fashion, forgetting elements that
the human system employs: temporal and contextual information, and cue com-
bination.
The work summarized in this abstract describes an approach for robust real-
time multiple face detection which combines different cues. The resulting ap-
proach achieves better detection rates for video stream processing and cheaper
processing costs than outstanding and public available face detection systems.
2 The Face Detection Approach

The approach briefly described makes use for the first detection or after a failure,
of two window shift detectors based on the general object detection framework
described in [6], which provide acceptable performance and processing rates.
These two brute force detectors, recently integrated in OpenCV, are the frontal
face detector described in that paper, and the local context based face detector
described in [3]. The last one achieves better recognition rates for low resolution
images and non frontal faces whenever the head and shoulders are visible.
The exclusive use of a monolithic approach based on the Viola framework has
the disadvantage of not using a main cue needed for video processing: temporal
coherence. Any face detected in a frame provides information which can be used
in the next frames to speed up the process. Therefore, for each detected face,
the system stores not only its position and size, but also its average color using
red, green normalized color space.
?
Work partially funded by research projects Univ. of Las Palmas de Gran Ca-
naria UNI2003/06 and Canary Islands Autonomous Government PI2003/160 and
PI2003/165.
Skin color based approaches for face detection have a lack of robustness for
different conditions. A well known problem is the absence of a general skin color
representation for any kind of light source and camera [5]. However, the skin color
extracted from the face previously detected by the frontal face Viola detector
can be used to estimate facial features position by means of the color blob, which
provides valuable information to detect eye positions for frontal faces [1].
Each face detected in a frame can be characterized by different features f =
hpos, size, red, green, leyepos , leyepattern , reyepos , reyepattern , f acepattern i. These
features direct different cues in the next frame which are applied opportunisti-
cally in an order based on the computational cost and the reliability.
– Eye tracking: A fast tracking algorithm [2] is applied in an area that sur-
rounds previously detected eyes, if available.
– Face detector: The Viola-Jones detector is applied in an area that covers the
previous detection [6].
– Local context face detector: If previous techniques fail, it is applied in an
area that includes the previous detection [3].
– Skin color: Skin color is searched in the window that contains the previous
detection, and the new sizes and positions coherently checked.
– Face tracking: If everything else fails, the prerecorded face pattern is searched
in an area that covers previous detection [2].
These techniques are applied until one of them finds the face, or the process
will be restarted using the Viola-Jones based detectors applied to the whole
image. Whenever a face is detected, the skin color is used for facial features
detection [1].
The approach extracts different features from each detected face in a frame,
therefore multiple face detection is considered. Additionally, a single or multi-
ple faces detected in consecutive frames are related according to their specific
features. During the video stream processing, the face detector gathers a set of
detection threads, IS = {dt1 , dt2 , ..., dtn }. A detection thread contains a set of
continuous detections, i.e. detections which take place in different frames but are
related by the system in terms of position, size and pattern matching techniques.
The Viola-Jones based detectors have some level of false detections. For that
reason a new detection thread is created only if the eyes are detected. The use of
the weakest cues, i.e. color and tracking, after a recent detection is reserved to
detections which are already considered part of a detection thread. In this way,
spurious detections do not launch cues which are not robust enough, in the sense
that they are not able to recover from a false face detection. The final results
is that for each detection thread, the face detector system provides a number
of facial samples, dtp = {x1 , ..., xmp }, which correspond to those detections for
which also the eyes were located.
3 Results
For still images the lower boundary is the combination of the Viola-Jones based
detectors [3, 6]. If both detectors return a face in the same position, it is preferred
the frontal face detection as it is more precise. For still images the added value
of the approach is the likely eye detection for almost frontal views and the
combination of two Viola-Jones based detectors, see Figure 1.
Fig. 1. Detection examples from some CMU database [4] samples. The color indicates
the technique used: green means that the eyes were detected, yellow means that they
were not detected, and red containing a yellow rectangle means that the local context
detector was used. The images have been scaled down to fit the extended abstract size.
The size of the images are originally 814 × 820, 256 × 256, 611 × 467 and 159 × 160
respectively in the first row and in the second 500 × 500, 539 × 734, 336 × 484 and
258 × 218. Obviously for still images there are no detections based on color or tracking.
The benefits of our approach are evidenced in video stream processing. The
results achieved processing a set of 70 desktop sequences containing more than
35000 images of different individuals allow, for typical webcam resolutions 320 ×
240, multiple face detection in real-time, an average of 20 fps. The approach
provides better recognition rates than the OpenCV implementation of the Viola
frontal face detector (3 percentage points greater), faster, and with the added
value of eye detection for almost frontal faces. The approach is also suitable for
sequences with resolution changes, see Figure 2.
The temporal coherence applied to video stream processing not only speeds
up the process but also allows to despise most false detections which appear
when a still image is processed, see some false detections in Figure 1. In at
least 10 of the sequences some detections were non face patterns, and they were
correctly not assigned to any detection thread as the eyes were not found and
their position, or their color and size were not coherent with any active detection
thread. Or in the worst case, a non face detection was associated to a detection
thread, but the system observed soon an incoherence and decided to remove the
detection thread and wait for a new one, i.e. a new eye pair detection.
As a conclusion, the approach developed provides good and real-time per-
formance for multiple face detection at different resolutions, making the system
suitable for further processing in the field of perceptual user interfaces.
Fig. 2. Frames extracted from a video stream with 720 × 576 resolution. The color has
the same meaning than in Figure 1, but observe that the last frame depicts a blue
rectangle which means that tracking was used.
References
1. M. Castrillón Santana, F.M. Hernández Tejera, and J. Cabrera Gámez. Encara: real-
time detection of frontal faces. In International Conference on Image Processing,
Barcelona, Spain, September 2003.
2. Cayetano Guerra Artal. Contribuciones al seguimiento visual precategórico. PhD
thesis, Universidad de Las Palmas de Gran Canaria, Octubre 2002.
3. Hannes Kruppa, Modesto Castrillón Santana, and Bernt Schiele. Fast and robust
face finding via local context. In Joint IEEE Internacional Workshop on Visual
Surveillance and Performance Evaluation of Tracking and Surveillance (VS-PETS),
October 2003.
4. Henry Schneiderman and Takeo Kanade. A statistical method for 3d object de-
tection applied to faces and cars. In IEEE Conference on Computer Vision and
Pattern Recognition, 2000.
5. Moritz Storring, Hans J. Andersen, and Erik Granum. Physics-based modelling of
human skin colour under mixed illuminants. Robotics and Autonomous Systems,
2001.
6. Paul Viola and Michael J. Jones. Rapid object detection using a boosted cascade
of simple features. In Computer Vision and Pattern Recognition, 2001.
7. Ming-Hsuan Yang, David Kriegman, and Narendra Ahuja. Detecting faces in
images: A survey. Transactions on Pattern Analysis and Machine Intelligence,
24(1):34–58, 2002.
Evolutionary Color Constancy Algorithm based on the
Gamut Mapping Paradigm
Cristian Munteanu1,2, Agostinho Rosa1, Manuel Galan2, Enrique Rubio Royo2
1 Instituto de Sistemas e Robotica, Instituto Superior Tecnico, Av. Rovisco Pais 1, Torre
Norte 6.21, 1049-001, Lisboa, Portugal
{cmunteanu, acrosa}@isr.ist.utl.pt
http://laseeb.isr.ist.utl.pt/
2 CICEI, Universidad de Las Palmas de Gran Canaria, Edificio Central del Parque Científi-
co y Tecnológico, Campus Universitario de Tafira, 35017 – Las Palmas, Spain

rubio@cicei.com
Abstract. In recent years, extensive work has been done to design algorithms
that strive to mimic the robust human vision system which is able the perceive
the true colors and discount the illuminant from a scene viewed under light hav-
ing different spectral compositions (the feature is called “color constancy”). We
propose a straightforward approach to the color constancy problem by employ-
ing an Interactive Genetic Algorithm [1] (e.g. a Genetic Algorithm [2], [3]
guided by the user) that optimizes a well known and robust variant of color
constancy algorithm called “gamut mapping” [4]. Results obtained on a set of
test images and comparison to various color constancy algorithms, show that
our method achieves a good color constancy behavior with no additional
knowledge required besides the image that is to be color-corrected, and with
minimal assumptions about the scene captured in the image.
1 Introduction
Color Constancy algorithms are fundamental both in Computer Vision whenever

recognition of objects in a scene is performed based on the color of the respective
objects, as well as for color correction of digital pictures with applications in medi-
cine, remote sensing, arts and media. When compared to the human visual system,
algorithms that perform Color Constancy fail short to achieve the same effectiveness
in recuperating the colors in scenes taken under different illuminants. This is due to
the fact that such algorithms make restrictive assumptions about the world from
which the image was taken. More-over, several algorithms require additional infor-
mation about the scene and the technical characteristics of the camera that captures
the image, knowledge that is not always readily available. We tackle the color con-
stancy problem by employing an Interactive Genetic Algorithm [1] (e.g. a Genetic
Algorithm guided by the user) that optimizes a well known and robust variant of
color constancy algorithm called “gamut mapping” [2]. Results are given on a set of
test images and comparison is made to various color constancy algorithms that proved
efficient and that pertain to the following categories: Gray World (GW) methods,
Gamut Mapping methods, and Neural Networks methods [2].
1.1 Gamut mapping as basis transformation for IGA-GM
The set of all possible responses due to known or expected surface reflectances, as
seen under a known, canonical illuminant (usually taken to be the daylight, or an
approximation of it), is a convex set, referred to as the canonical gamut. Similarly, the
set of responses due to an unknown illuminant is also a convex set. Illumination
change is assumed to be modeled by a simple diagonal transformation according to
the research of von Kries which was later confirmed to hold as a reasonable assump-
tion in [3]. It was shown that under the von Kries adaptation assumption the two
gamuts are within a diagonal transformation of each other. Due to the fact that GM
proposes a simple diagonal transformation to each pixel in the image, and because it
has been shown that under not very restrictive assumptions such diagonal transforma-
tions suffice for good color constancy [3], in this paper we adopt the GM method as
the basic color constancy transformation applied to correct a given image. The main
drawbacks of the classical GM methods are: a) they use extensive knowledge about
the camera used to take the picture; b) they use a priori information about the world
in which the image was taken in order to construct the gamuts.
2 IGA-GM. Algorithm presentation
In the present paper we employ an Interactive Genetic Algorithm (IGA) in which a

human evaluator iteratively gives a subjective score for each result of a diagonal
mapping (corrected image), until the best corrected image (and implicitly the best
diagonal mapping) is discovered. Thus IGA-GM adapts to the “best” gamut mapping
(diagonal transformation) according to human subjective criteria. This is done with-
out making use of any a priori knowledge about the surrounding world conditions, or
the camera. The Genetic Algorithm [4] is given in pseudocode in Fig.1. Each individ-
ual x i in the population (chromosome) codes a diagonal transformation which is next
applied to each pixel in the original (input) image, to give the corrected (output) im-
age:
∀i, j : ( Rpixel( i , j ) , G pixel( i , j ) , B pixel( i , j ) ) = ( x i ,1 ⋅ R pixel( i , j ) , x i , 2 ⋅ G pixel( i , j ) , x i , 3 ⋅ B pixel( i , j ) )

corrected corrected corrected original original original
(1)
How well a chromosome performs (that is, how good the diagonal mapping is in
recuperating the original colors and how well it corrects the input image) is judged
mainly by the human evaluator who looks at the corrected image and gives a numeric
score, called fitness value: f ( xi ) ranging from 0 (worst appearing image) to 10 (best
appearing image). The user shouldn’t evaluate all images corrected by each chromo-
some in the population, in each generation of the algorithm, because such a process
would become too tedious (there would be a lot of images to evaluate). Many of the
chromosomes to be evaluated are actually given a fitness value automatically, using a
clustering algorithm over fitness values previously allocated by the user.
t:=0
initialise P(t)
evaluate_1 P(t)
while (terminate(P(t)) ≠ true) do
{
P’(t) := select_tournament (P(t) | q)
P’’(t):= crossover(P’(t) | Pc)
P’’’(t):= mutate (P’’(t) | Pm)
P(t+1):= elitist(P’’’(t) ∪ P(t) | K)
evaluate_2 P(t+1)
t := t+1
}
endwhile
Fig. 1. Pseudocode of IGA-GM. t represents the time/generation index; P(t) is the population at
generation t, P’(t) is the population at generation t after selection, P’’(t) is the population at
generation t after selection and crossover, and P’’’(t) is the population at generation t after
selection, crossover and mutation. K is the number of elites.
3 Experimental results
The experimental part comprises two main sections: a) firstly we perform a com-
parison between IGA-GM and other Color Constancy algorithms for a single un-
known illuminant (e.g. different than canonical or daylight illuminant). b) secondly,
we check the effectiveness of IGA-GM on a set of scenes taken under two quite dif-
ferent than daylight illuminants.
Fig. 2. Comparison to Color Constancy algorithms (Detergent image). Upper-row from left to
right: original image (to be color-corrected), target image (image under canonical illuminant),
GW, Retinex; Lower-row from left to right: NN, GM, IGA-GM.
Table 2. RMS error calculated for Detergent image on pairs between target image (T) and
Color Constancy corrected image. The image that is not corrected (the original - input image)
is denoted as O.
(T, O) (T,GW) (T, Retinex) (T,NN) (T, GM) (T, IGA-GM)
RMS(R,G,B) 2.34e+04 18.02 49.13 13.38 22,30 16.16
RMS(r,g) 16.42 12.88 16.27 7.78 21.74 7.99
Ball image
Fig.3 IGA-GM solutions on different illuminants. Figure is divided in horizontal bands for
each image/scene. For each band we have: first column first line: target image; second line first
column: image under “solux_4700+3202”; second line second column: IGA-GM solution;
second line third column: “IGA-GM solution + intensity correction”; third line first column:
image under “syl_wwf”; third line second column: IGA-GM solution; third line third column:
IGA-GM solution + intensity correction.
References
1. Takagi, H.: Interactive Evolutionary Computation: Fusion of the Capabilities of EC

Optimization and Human Evaluation. Proceedings of the IEEE, vol.89, (2001) 1275-1296
2. Barnard, K., Cardei, V., Funt, B. V.: A Comparison of Computational Color Constancy
Algorithms-Part I: Methodology and Experiments with Synthesized Data. IEEE Trans. on
Image Processing, vol. 11, no. 9, (2002) 972-984
3. Finlayson, G., Drew, M. S., Funt, B.V.: Diagonal transforms suffice for color constancy.
Proc. IEEE Int. Conf. on Computer Vision, (1993) 164-171
4. Michalewicz, Z.: Genetic Algorithms + Data Structures = Evolution Programs. 3rd edn.
Springer-Verlag, Berlin Heidelberg New York (1996)
Vision based Automatic Occupant Classification and
Pose Recognition for Smart Airbag Deployment
Min-Soo Jang1, Yong-Guk Kim2, Sang-Jun Kim1, Jeong-Eom Lee1,

Seok-Joo Lee3, and Gwi-Tae Park1*
1Dept. of Electrical Engineering, Korea University, Seoul, Korea
gtpark@korea.ac.kr
2 School of Computer Engineering, Sejong University, Seoul, Korea
ykim@sejong.ac.kr
3 Hyundai Autonet Co. Ltd., Korea
Extended Abstract
During the last decade, the safety of occupants in the cars has been increased by im-
proving the airbag system, which can prevent life-threatening head (or chest) injuries
by avoiding direct contact to the dashboard during the accident. However, it is also
well known that the airbag deployment to the child (or infant) occupant could lead to
an injury [6]. To reduce the risk caused by airbag, it is important to design a smart
airbag that could control triggering and intensity of its deployment. In this paper, we
present a vision based automatic system, consisting of an occupant classification
system and an occupant’ pose recognition system as illustrated in Fig. 1 (a), by which
we aim to control whether the airbag should be triggered or not, and how strongly it
should be deployed when it is triggered.
(a) (b)
Fig. 1. The block diagram of the airbag control system (a) and the head tracking for the pose
recognition (b) - the solid circle indicates the head position of an occupant and the numbers
show the areas where the occupant can take a pose.
The occupant classification system can categorize the occupant into four differ-
ent classes such as the 5th%tile female (or adult), 6 year-old child, infant seat, and
empty seat as shown in Fig. 2. From a pair of stereo images acquired with a stereo-
camera, the disparity map is extracted by using a fast SAD (Sum of the Absolute
Difference) algorithm and a cubic model is used for background elimination [1]. And
the down-sampled image of the disparity map is used as an input for the SVM (Sup-
port Vector Machine) classifier [2]. Since the disparity maps of the 5th%tile female
and 6 year-old child are similar compared to other cases, the SVM classifier consists
of two stages: in the first stage, it classifies the occupant into three classes, such as
the 5th%tile female or 6 year-old child as the first class, the infant seat as the second
class, and the empty seat as the third class; in the second stage, the first class case is
classified as either the 5th%tile female or 6 year-old child. According to the classifica-
tion outcome, the system will deploy the airbag only if the occupant is classified as
the 5th%tile female, otherwise it will not be deployed.
Fig. 2. The occupant classification system
For the pose recognition system, since we assume that the head of the occupant
can take the diverse poses, the passenger area is divided into 7 areas. And the system
is to recognize which area the occupant’ head belongs to, as shown in Fig. 1 (b). To
make the system robust, two algorithms are working in parallel as shown in Fig. 3.
First, we try to detect the occupant’ head in the given image by using a head contour
model [3] extracted from the accumulative difference images by combining a SVM
classifier that was trained using different sizes of head contours as illustrated in Fig. 4
(a) where each head contour model is made up different number of feature points, 97,
81 and 65, respectively. When the head detection succeeds, the system initializes
tracking parameters and repeats the sequence; otherwise, it checks whether the track-
ing parameters are initialized or not in the previous stage. For the head tracking, it
tracks the head by using an inverse compositional algorithm according to Lucas-
Kanade algorithm [4, 5]. When it fails, it counts as an error and returns to the head
detection stage. If the number of error becomes larger than the permitted number of
errors, the system halts and sends a warning signal.
Fig. 3. Flow chart of the pose recognition system
Fig. 4 (b) illustrates the detail schematic diagram of the head tracking. In the
initialization part, a model image is extracted from the given image and then it initial-
izes the tracking parameters, such as warping and motion parameters. In the head
tracking part, a warped image is made using the tracking parameters and it updates
the tracking parameters to minimize the sum of squared error between the model
image and the warped image. Such step is iterated until the tracking parameters con-
verge. Notice that the inverse compositional algorithm is fast enough to track the
head in the real-time base.
(a) (b)
Fig. 4. The head contour model with three different sizes of head (a) and the block diagram of
the head tracking algorithm (b).
The performance of occupant classification system was evaluated with our ste-
reo image database acquired within the experimental car. The stereo images totaled
802 and the resolution of each image was 320x240. We used down-sampled images
(resolution: 80x60) for reducing dimension of the data. Table 1 (a) shows average
results of occupant classification. The correction rate was 96.57% and 96.43% for the
case of RBF kernel and polynomial kernel, respectively. And, in case of classifying
the occupant at a time by using SVM classifier, the correction rate was 91.47%. Table
1 (b) shows results of occupant’ pose recognition. The head tracking experiment was
accomplished by using two image sequences which consist of 608 and 411 frames,
respectively. A success case was counted if any one of the two algorithms recognize
the occupant’ head as shown in Fig. 1 (b). Results show that the performance of the
system is satisfactory, suggesting that the vision based airbag control has a potential
in real applications. We plan to carry out further study to solve the occlusion problem
caused by thick coat, arm movement or hat.
Table 1. Occupant classification rate for each kernel (a) and

pose recognition result for two image sequences (b)
(a) (b)
Kernel Classification Recogni-

Row data
type rate tion rate
Polynomial 96.57% Sequence set I 83.11%
RBF 96.43% Sequence set II 83.42%
References
1. H. G. Lee et al., Occupant Classification for Smart Airbag Using Stereovision and Support
Vector Machines, Springer Lecture Notes in Computer Science, Vol. 3173, pp. 642-647,
2004.
2. V. N. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag New York, Inc.,
New York, NY, 1995.
3. M. W. Lee, Cohen I., S. K. Jung, Particle filter with analytical inference for human body
tracking, Motion and Video Computing, Workshop on., pp. 159-165, 2002.
4. B. Lucas, T. Kanade, An Iterative Image Registration Technique with an Application to
Stereovision, Proceedings of IJCAI 81, pp. 674-679, 1981.
5. S. Baker, I. Matthews, Lucas-Kanade 20 Years On: A Unifying Framework: Part 1. Techni-
cal Report, CMU-RI-TR-02-16, Carnegie Mellon University Robotics Institute, 2002
6. http://www.nhtsa.dot.gov
Wiener-type neuronal model with refractoriness
V. Giorno1 , A.G. Nobile1 and L.M. Ricciardi2

1
Dipartimento di Matematica e Informatica, Università di Salerno, Via Ponte don
Melillo, Fisciano (SA), Italy, {giorno,nobile}@unisa.it
2
Dipartimento di Matematica e Applicazioni, Università di Napoli Federico II, Via
Cintia, Napoli, Italy, luigi.ricciardi@unina.it
Extended Abstract
The extension of the Wiener neuronal model [3] in a way to include a con-
stant refractoriness time is the object of the present article. We shall denote by
{X(t), t ≥ 0} the Wiener process defined in I = (−∞, +∞) and characterized
by drift A1 (x) = µ and infinitesimal variance A2 (x) = σ 2 (µ > 0, σ > 0). To this
purpose, we assume that after the firing a period of refractoriness of constant
duration 1/ξ (ξ > 0) occurs, during which either the neuron is completely unable
to respond or it only partially responds to the incoming stimulations (cf., for in-
stance, [1], [2], [5]). We construct the return process {Z(t), t ≥ 0} in (−∞, S)
as follows (cf. [4]). Starting at a fixed state η ∈ (−∞, S) at time zero, a firing
takes place when X(t) attains for the first time a boundary S representing the
neuron’s threshold, after which the neuron is unable to fire again for the time
duration 1/ξ. At the end of the period of refractoriness, Z(t) is instantaneously
reset to η. The subsequent evolution of the process then goes on as described by
X(t), until the boundary S is again reached. A new firing then occurs, followed
by the period of refractoriness, and so on.
Hereafter, we shall provide a description of the random process {M (t), t ≥ 0}
representing the number of firings released by the neuron up to time t and
a characterization of the interspike pdf. To this purpose, for η ∈ (r1 , S), the
probability of having k firings up to time t is proved to be:
t−(k−1)/ξ
k−1
qk (t|η) := P {M (t) = k|Z(0) = η} = H t − g(k S, ϑ | k η) dϑ
ξ 0
t−k/ξ
k
−H t − g (k + 1) S, ϑ | (k + 1) η dϑ
ξ 0
where H(t) denotes the Heaviside unit step function and

z−y (z − y − µ t)2
g(z, t|y) = √ exp − (z > y)
σ 2 π t3 2 σ2 t
is the first passage time (FPT) pdf of the considered Wiener process. The first
two moments of the number of firings released by the neuron up to time t are
then obtained:
t ξ t−k/ξ

E M (t) | η := k qk (t|η) = g (k + 1) S, ϑ | (k + 1) η dϑ,
k≥1 k=0 0
t ξ
2 t−k/ξ
E [M (t)]2 |η := k qk (t|η) = (2 k + 1) g (k + 1) S, ϑ | (k + 1) η dϑ,
k≥1 k=0 0
where x denotes the largest integer less than or equal to x.

Let now I0 , I1 , I2 , . . . denote the random variables describing the neuronal
output, with I0 representing the time of occurrence of the first firing and Ik (k =
1, 2, . . .) the duration of the time interval elapsing between k-th and (k + 1)-th
firing. Furthermore, let γk (t) denote the pdf’s of Ik (k = 0, 1, . . .). Therefore, I0
identifies with the FPT through the threshold S starting at initial state X(0) =
η < S, so that γ0 (t) ≡ g(S, t|η). The interspike intervals I1 , I2 , . . . are iid random
variables having pdf’s:

 0, t ≤ 1/ξ
γk (t) = (k = 1, 2, . . .).

g S, t − 1/ξ | η , t > 1/ξ
Hence, the mean and variance of the interspike intervals are given by:

 S−η

 , k=0
 µ σ 2 (S − η)
E(Ik ) = Var(Ik ) = , k = 0, 1, . . . .

 1 S−η µ3

 + , k = 1, 2, . . .
ξ µ
Extension to random refractoriness in neuronal diffusion models will be the
object of future works.
References
1. Buonocore, A., Giorno, V., Nobile, A.G., Ricciardi, L.M.: A neuronal modeling
paradigm in the presence of refractoriness. BioSystems 67 (2002) 35-43
2. Esposito, G., Giorno, V., Nobile, A.G., Ricciardi, L.M., Valerio, C.: Interspike anal-
ysis for a single neuron’s activity in presence of refractoriness. In: Trappl, R. (ed.):
Cybernetics and Systems. Vol. 1. Austrian Society for Cybernetics Studies, Vienna
(2004) 199-204
3. Gerstein, G.L., Mandelbrot, B.: Random walk models for the spike activity of a
single neuron. Biophys. J. 4 (1964) 41-68
4. Ricciardi, L.M., Di Crescenzo, A., Giorno, V., Nobile, A.G.: An outline of theoretical
and algorithmic approaches to first passage time problems with applications to
biological modeling. Math. Japonica 50 No. 2 (1999) 247-322
5. Ricciardi, L.M., Esposito, G., Giorno, V., Valerio, C.: Modeling Neuronal Firing
in the Presence of Refractoriness. In: Mira, J., Alvarez, J.R. (eds): Computational
Methods in Neural Modeling. IWANN 2003. Lecture Notes in Computer Sciences,
Vol. 2686. Springer-Verlag (2003) 1-8
On Myosin II dynamics: from a pulsating
ratchet to a washboard potential?
A. Buonocore, L. Caputo, E. Pirozzi and L.M. Ricciardi

Cintia, Napoli, Italy,
{aniello.buonocore,enrica.pirozzi,luigi.ricciardi}@unina.it
luigia.caputo@dma.unina.it
Abstract
A Brownian motor for a Myosin II (“myosin”, in the sequel) head sliding
along an actin filament is designed. Various models based on the overdamped
equation s
U 0 [x(t), f(t)] 2kB T
ẋ(t) = − + Λ(t) (1)
ρ ρ
have often been proposed (see, for instance, [1], [2], [3]). Eq. (1) mimics the
myosin head as a Brownian particle subject to the impacts by the surrounding
molecules. In (1), Λ(t) is a standard Gaussian white noise, f(t) a random pro-
cess whose nature must be specified, and ρ the myosin friction drag coefficient.
U 0 [x(t), f(t)], the space derivative of a periodic potential U = U (x, t), stands
for the opposite of the global force arising from all deterministic interactions of
myosin head with the actin filament. The specification of f(t) marks the differ-
ence among the proposed models. In particular, in [1] the potential is assumed to
be an asymmetric saw-tooth function whose height switches between 2 different
values at Poisson distributed instants. In [2], U is a single saw-tooth potential,
the positions of whose minima change according to a Poisson law. Here, we shall
refer to this second formulation, by adding the assumptions that the motor step
equals the distance L between successive actin monomers (roughly 5.5 nm). Fur-
thermore, we assume that two 2L-periodic potentials, U1 (x) and U2 (x), alternate
according to a Poisson process, each potential being spatially shifted by L with
respect to the other. We denote by as the constant Poisson rate determining the
transition Us (x) → U3−s(x), s ∈ {1, 2}, of the particle’s position. The dynamics,
governed by s
Us0 (x) 2kB T
ẋ(t) = − + Λ(t) (2)
ρ ρ
is sketched in Fig. 1. Along the lines of [1], the transition pdf ps (x) and the corre-
sponding probability current Js (x) are determined within the reduced dynamics
in the stationary regime. It is shown that via certain averages with respect to
p(x) = p1 (x) + p2 (x), some features of the motor can be quantitatively deter-
mined. In particular, (i) the numerical solution of a fixed point equation (as
?
(Workshop “Computation and Simulation in Modelling Biological Systems”)
in [4]) leads to the stall force for the motor; (ii) by exploiting an idea of [5], the
derivative of the effective driving potential is calculated, and thus the intensity of
the force generated by the motor is determined; and (iii) the engine power with
reference to the reactions of 1 chemical cycle is calculated. Having determined
the force generated by the motor, via a Boltzmann equation the conditional
probability p of a forward step is estimated. Then, the mean duration µ between
pairs of successive steps is obtained in terms of p and J(x) = J1 (x)+J2 (x). Such
a value can be compared with that obtained via
Z 2L Z x
ρ V (x) V (y)
µ̂ = p dx exp dy exp − (3)
kB T 0 kB T x−2L kB T
providing the analogous quantity for a model based on washboard-type potential
(see [6]). The agreement between µ and µ̂ is pointed out by setting in (3) V (x) =
U (x) + Fe, where U (x) denotes the effective driving potential and Fe represents
a constant externally applied force.
Chemical state transitions

s s
as
3-s 3-s
Particles' displacement
. X
. X
L L
2L 2L
Fig. 1. Sketch of the Brownian motor. On top state transitions induced by a

chemical reaction are indicated. In the bottom, the corresponding activated
potentials (thick lines), and the previous potentials (dotted lines) are drawn.
Possible positions of two particles (full and empty circles), as a consequence
of potential changes, are indicated.
References
1. Astumian, R.D., Bier, M.: Fluctuation Driven Ratchets: Molecular Motors. Physical
Review Letters, Vol. 72, (1994) 1766–1769
2. Chauwin, J.F., Ajdari, A., Prost, J.: Force-Free Motion in Asymmetric Structures:
a Mechanism without Diffusive Steps. Europhysics Letters, Vol. 27, (1994) 421–426
3. Prost, J., Chauwin, J.F., Peliti, L., Ajdari, A.: Asymmetric pumping of particles.
Physical Review Letters, Vol. 72, (1994) 2652–2655
4. Chiu, C., Peskin, C.S.: Numerical analysis of the flux and force generated by a
correlation ratchet mechanism. Mathematical Biology, Vol. 44, (2002) 479–501
5. Wang, H., Oster, G.: Ratchets, power strokes, and molecular motors. Applied
Physics A, Vol. 75, (2002) 315–323
6. Lindner, B., Kostur, M., Schimanky-geier, L.: Optimal diffusive transport in a tilted
periodic potential. Fluctuation and Noise Letters, Vol. 1, (2001) R25–R39
Feedback effects in Stein’s coupled neurons
A. Di Crescenzo1 , B. Martinucci2 and E. Pirozzi2

1
Dipartimento di Matematica e Informatica, Università di Salerno
Via Ponte don Melillo, I-84084 Fisciano (SA), Italy
adicrescenzo@unisa.it
2
Dipartimento di Matematica e Applicazioni, Università di Napoli Federico II
Via Cintia, I-80126 Napoli, Italy
barbara.martinucci@dma.unina.it, enrica.pirozzi@unina.it
Extended Abstract
The dynamics of a pair of Stein’s neuronal units serially connected has been
recently analyzed ([4], [5]), under the assumption that they are subject to ex-
citatory and inhibitory stimuli with constant or with alternating rates. Making
use of a simulation procedure designed by some of us, various quantitative re-
sults have been obtained. In particular, it has been disclosed the existence of an
optimal value for the amplitude γ of the jumps of the membrane potential of the
second unit (the “receiving neuron”) attained as effect of the firing of the first
unit (the “sending neuron”). We also obtained results concerning the reaction
times, defined as the random times elapsing between a firing of the sending neu-
ron and the subsequent firing of the receiving neuron. The reaction times have
been analyzed in [5] under the assumption of constant inhibitory rate and for
both constant and alternating excitatory rate, with emphasis on the effects of
various choices of γ and of the amplitude of the refractory period.
The model considered in [4] and [5] is re-considered here under the new
assumption that a complete interaction is possible between the two neurons in
the sense that the firing of each neuron causes a jump of constant amplitude of
the membrane potential of the other neuron. The positive or negative value of
the jumps reflects the excitatory and inhibitory nature of firing neuron.
A first step of our study consists of the analysis of the interspike intervals
(ISI’s) of the two neurons by analyzing their histograms and empirical distri-
bution functions. Our results are then compared with those of [4], concerning a
serial connection between the two neurons.
To look at feedback effects in the firing activity of the two neurons we have
evaluated the relative entropy of the ISI’s distributions. Moreover, recalling the
statistical analysis performed by Elson et al. [6], we have focused our attention
on the reaction times and on ISI’s of the neurons in order to disclose their
synchronization. Some comparisons have been performed on the basis of the
values of the following index:
(2) (1) (1) (1)
Ri = (ti − tj )/(tj+1 − tj ),
(1) (2)
where tj is the firing time of the first neuron preceding the firing time ti of
the second one. Such an index measures the delay of the spike emission of the
second neuron ensuing the firing of the first neuron.
The synchronization of the firing activity of the neurons has been analyzed in
the special case in which one of the coupled neurons is excitatory and the other
is inhibitory. The relevance of the analysis of firing activity of pairs of neurons
under various mutual interactive paradigms is of interest since, for instance, it
is known that the cerebellar cortex includes pairs of coupled neurons, which has
motivated previous studies. (See, for instance, [2] and [3]).
Finally, we point out that our analysis of the synchronization of the firing
activity of the neurons has been performed also by studying the cross-correlation
function (see [1]). The auto-correlation function evaluated with reference to the
firing times of the two neurons has allowed us to disclose particular patterns of
spike trains from single neurons, including oscillating behaviors.
References
1. Borisyuk, R.: Oscillatory activity in the neural network of spiking elements. BioSys-
tems 67 (2002) 3–16
2. Burkitt, A.N.: Balanced neurons: analysis of leaky integrate-and-fire neurons with
reversal potential. Biological Cybernetics 85 (2001) 247–255
3. Cartling, B.: Control of computational dynamics of coupled integrate-and-fire neu-
rons. Biological Cybernetics 76 (1997) 383–395
4. Di Crescenzo, A., Martinucci, B., Pirozzi, E., Ricciardi L.M.: On the interaction
between two Stein’s neuronal units. In: Trappl, R. (ed.): Cybernetics and Systems
2004, Vol. 1. Austrian Society for Cybernetics Studies, Vienna (2004) 205–210
5. Di Crescenzo, A., Martinucci, B., Pirozzi E.: On the dynamics of a pair of coupled
neurons subject to alternating input rates. BioSystems (2004) (to appear)
6. Elson, R.C., Selverston, A.I., Huerta, R., Rulkov, N.F., Rabinovich, M.I., Abarbanel,
H.D.I.: Synchronous Behavior of Two Coupled Biological Neurons. Physical Review
Letters 81, N. 25 (1998) 5692–5695
Upcrossing first passage times for correlated
Gaussian processes
V. Giorno1, A.G. Nobile1 and E. Pirozzi2

1
Dipartimento di Matematica e Informatica, Università di Salerno, Via Ponte don
Melillo, Fisciano (SA), Italy, {giorno,nobile}@unisa.it
2
Cintia, Napoli, Italy, enrica.pirozzi@unina.it
Extended Abstract
The first passage time (FPT) probability density function (pdf) of a stochas-
tic process through certain boundaries is used to model the first time when
the neuronal membrane potential crosses the threshold value thus generating
an action potential (see, for example, [1] and [5]). Along the line of previous
mathematical and computational approaches pursued to solve the so-called FPT
problem ([4],[2],[3]), here we investigate the asymptotic behavior of the FPT pdf
for a class of stationary Gaussian processes for increasing correlation times.
Let X(t) be a non-singular stationary Gaussian process with zero mean and
covariance E[X(t)X(τ )] = %(t − τ ) such that %(0) = 1, %̇(0) = 0 and %̈(0) < 0;
then, Ẋ(t), the derivative of X(t) with respect to t, exists in the mean-square
sense. Since X(t) is stationary, the correlation time is as ([5]):
Z +∞
τc := %(ϑ) dϑ.
0
We shall focus on the FPT upcrossing problem that we define as an FPT prob-
lem to a boundary S(t) ∈ C 1[0, +∞) for the subset of sample paths of X(t)
originating at time zero at a state X0 . The state is in turn viewed as a random
variable with pdf


 f(x0 )
 , x0 < S(0) − ε
P {X(0) < S(0) − ε}
γε (x0 ) ≡



0, x0 ≥ S(0) − ε,
where ε > 0 is a fixed real number and f(x0 ) denotes the normal pdf of X(0).
Then,
(ε)
TX0 := inf {t : X(t) > S(t)},
t≥0
is the ε-upcrossing FPT of X(t) through S(t) and its pdf is given by
Z S(0)−ε
∂ (ε)
gu(ε) (t) = P (TX0 < t) = g(t|x0) γε (x0 ) dx0 (t ≥ 0),
∂t −∞
where g(t|x0) denotes the FPT pdf of X(t) through S(t) conditional on X(0) =
x0 < S(0).
By using a straightforward variant of a method proposed by Ricciardi and
Sato ([6]) for the determination of the conditional FPT density, in ([2]) we have
obtained a series expansion for the upcrossing FPT pdf. However, the evaluation
of the partial sums is effectively impossible due to the outrageous complexity of
the involved functions and of their integrals. Nevertheless, if the correlation
function of X(t) satisfies
%̇(t) 1
lim %(t) = 1, lim %̇(t) = lim %̈(t) = 0, lim = − (1)
τc →+∞ τc →+∞ τc →+∞ τc →+∞ 1 − %2 (t) t
and if S(t) is a monotonically decreasing function, then the following approxi-
mate behavior holds as τc → +∞:
gu(ε) (t) ' −Ṡ(t) γε [S(t)], if S(t) < S(0) − ε, and gu(ε) (t) ' 0, otherwise.
An example of a process of concrete interest for applications that satisfies
(1) is characterized by the following covariance function:
1
%(t) = e−α|t|{cos(αω|t|) + sin(αω|t|)}, (t ∈ R) (2)
ω

where α > 0, ω ∈ R and τc = 2/ α(1 + ω2 ) .
In the present paper we indicate the practical usefulness of the above approx-
imation by means of reliable statistical estimations on numerically simulated
processes having covariance (2).
References
1. Di Nardo, E., Nobile, A.G., Pirozzi, E., Ricciardi, L.M.: On a non-Markov neuronal
model and its approximation. BioSystems, Vol. 48, (1998) 29–35
2. Di Nardo E., Nobile A.G., Pirozzi E., Ricciardi L.M.: Gaussian Processes and Neural
Modeling: an Asymptotic Analysis. In Cybernetics and Systems 2002, Vol. 1 (Trappl,
R., ed.), Proceedings of the 16th European Meeting on Cybernetics and Systems
Research, Vol I, Austrian Society for Cybernetics Studies, Vienna, ISBN 3-85206-
160-1, (2002) 313-318
3. Di Nardo E., Nobile A.G., Pirozzi E., Ricciardi L.M.: Computational Methods for
the Evaluation of Neuron’s Firing Densities. In Computer Aided Systems Theory
- EUROCAST 2003, (R. Moreno-Diaz, F. Pichler), Lecture Notes in Computer
Science 2809, Springer-Verlag Berlin Heidelberg New York, ISSN 0302-9743, ISBN
3-540-20221-8,(2003) 394-403
4. Giorno V., Nobile A.G., Pirozzi E., Ricciardi, L.M.: Non-stationary Gauss-Markov
Processes and Neuronal Modeling. In Cybernetics and Systems 2004 (Trappl R.,
ed.). EMCSR 2004 Austrian Society for Cybernetics Studies, Vienna, ISBN 3 85206
169 5, (2004) 211-215
5. Kostyukov, A. I., Ivanov,Yu.N., Kryzhanovsky, M.V.: Probability of Neuronal Spike
Initiation as a Curve-Crossing Problem for Gaussian Stochastic Processes. Biological
Cybernetics, Vol. 39, (1981) 157-163
6. Ricciardi, L.M., Sato, S.: On the evaluation of first–passage–time densities for Gaus-
sian processes. Signal Processing, Vol. 11, (1986) 339–357
Convergence of Iterations
Paul Cull
Oregon State University Corvallis

Oregon, 97331,USA
Extended Abstract
Convergence is a central problem in both computer science and in population
biology.
Will a program terminate?
Will a population go to an equilibrium?
In general these questions are quite difficult – even unsolvable.
In this talk we will concentrate on very simple iterations of the form
xt+1 = f (xt )
where each xt is simply a real number and f (x) is a reasonable real function.
Such one-dimensional iterations have special properties not shared by higher
dimensional iterations. For example, if f (x) is continuous and bounded, with
exactly one fixed point and no cycle of period 2, then the iteration will always
approach the fixed point.
For such a system, we say that an iteration is “globally stable” if it approaches
the fixed point for all starting points, and that the iteration is “locally stable” if
it approaches the fixed point for all initial points near enough to the fixed point.
Local stability can be checked by a simple derivative test, but showing global
stability requires more effort.
Population biologists have used one-dimensional iterations as simple models
of population growth. When they have shown local stability, they have assumed
global stability and have never run into any difficulty. Over the years a variety of
mathematical techniques have been used to show that specific simple population
models exhibit this
local stability implies global stability
phenomenon. But, no single method could prove this result for all the commonly
used population models.
We will show that there is a simple method which assures global stability.
Our method uses bounding of f (x) by a ratio of two linear functions. We call
this bounding “enveloping” and we show that
enveloping implies global stability.
Returning to the population models, we show that for these models local sta-
bility implies enveloping and hence global stability. For several models a single
enveloping function can be used for all parameter values, and we give a Tay-
lor series condition which implies this enveloping. For other models, different
enveloping functions are needed for different parameter combinations.
Although enveloping is a sufficient condition for convergence, we show by
example that enveloping is not necessary.
Finally we show that
enveloping implies global stability
even when f (x) is a discontinuous multi-function.
References
1. P. Cull. Global Stability of Population Models. Bulletin of Mathematical Biology,
43:47–58, 1981.
2. P. Cull. Local andGlobal Stability for Population Models. Biological Cybernetics,
54:141–149, 1986.
3. P. Cull. Local and Global Stability of Discrete One-dimensional Population Models.
In L. M. Ricciardi, editor, Biomathematics and Related Computational Problems,
pages 271–278. Kluwer, Dordrecht, 1988.
4. P. Cull. Stability of Discrete One-dimensional Population Models. Bulletin of
Mathematical Biology, 50(1):67–75, 1988.
5. P. Cull. Linear Fractionals - Simple Models with Chaotic-like Behavior. In D. M.
Dubois, editor, Computing Anticipatory Systems:CASYS 2001 - Fifth Interna-
tional Conference, Conference Proceedings 627, pages 170–181. American Institue
of Physics, Woodbury, N.Y., 2002.
6. P. Cull and J. Chaffee. Stability in discrete population models. In D. M. Dubois,
editor, Computing Anticipatory Systems: CASYS’99, pages 263–275. Conference
Proceedings 517, American Institute of Physics, Woodbury, NY, 2000.
7. P. Cull and J. Chaffee. Stability in simple population models. In R. Traappl, editor,
Cybernetics and Systems 2000, pages 289–294. Austrian Society for Cybernetics
Studies, 2000.
8. Paul Cull. Stability in one-dimensional models. Scientiae Mathematicae Japonicae,
58:349–357, 2003.
9. R. Devaney. An Introduction to Chaotic Dynamical Systems. Benjamin, Redwood
City, 1986.
10. M.E. Fisher, B.S. Goh, and T.L. Vincent. Some Stability Conditions for Discrete-
time Single Species Models. Bulletin of Mathematical Biology, 41:861–875, 1979.
11. M.P. Hassel. Density Dependence in Single Species Populations. Journal of Animal
Ecology, 44:283–296, 1974.
12. J.P. LaSalle. The Stability of Dynamical Systems. SIAM, Philadelphia, 1976.
13. T-Y. Li and J. Yorke. Period Three Implies Chaos. American Mathematical
Monthly, 82:985–992, 1975.
14. R.M. May. Simple Mathematical Models with Very Complicated Dynamics. Na-
ture, 261:459–467, 1976.
15. P.A.P. Moran. Some Remarks on Animal Population Dynamics. Biometrics, 6:250–
258, 1950.
16. A. Nobile, L.M. Ricciardi, and L. Sacerdote. On Gompertz Growth Model and
Related Difference Equations. Biological Cybernetics, 42:221–229, 1982.
17. C.J. Pennycuick, R.M. Compton, and L. Beckingham. A Computer Model for
Simulating theGrowth of a Population, or of Two Interacting Populations. Journal
of Theoretical Biology, 18:316–329, 1968.
18. W.E. Ricker. Stock and Recruitment. Journal of the Fisheries Research Board of
Canada, 11:559–623, 1954.
19. G. Rosenkranz. On Global Stability of Discrete Population Models. Mathematical
Biosciences, 64:227–231, 1983.
20. A. Sarkovskii. Coexistence of Cycles of a Continuous Map of a Line to Itself. Ukr.
Mat. Z., 16:61–71, 1964.
21. D. Singer. Stable Orbits and Bifurcation of Maps of the Interval. SIAM Journal
on Applied Mathematics, 35(2):260–267, Sept 1978.
22. S. Utida. Population Fluctuation, an Experimental and Theoretical Approach.
Cold Spring Harbor Symposium on Quantitative Biology, 22:139–151, 1957.
Semiautomatic Snake-Based Segmentation of
Solid Breast Nodules on Ultrasonography
Miguel Alemán-Flores1 , Patricia Alemán-Flores2 , Luis Álvarez-León1 , M. Belén

Esteban Sánchez1 , Rafael Fuentes-Pavón2 , and José M. Santana-Montesdeoca2
1
Departamento de Informática y Sistemas
Universidad de Las Palmas de Gran Canaria, 35017, Las Palmas, Spain
2
Sección de Ecografía, Servicio de Radiodiagnóstico
Hospital Universitario Insular de Gran Canaria, 35016, Las Palmas, Spain
Email: maleman@dis.ulpgc.es
Extended Abstract
Radiologists have described a series of criteria which help deciding whether a

solid breast nodule is malignant or benignant by analyzing ultrasound images [1].
These factors involve a precise visual examination of the shape of the nodule and
its contour. Thus, a measurement of how ellipsoid the nodule is, the extraction
of the ramifications, or the location of microlobulations, angular margins and
spiculations require the analysis of the whole contour or certain parts of it.
For this reason, it is very important to obtain an accurate delimitation of the
nodule boundary. Computer Vision techniques provide some methods for the
extraction of the contours, but, due to the special characteristics of ultrasound
images, they must be adapted to obtain satisfactory results. This paper presents
a new approach for the segmentation of breast tumors using active contours,
combined with other techniques, such as the truncated median, the structure
tensor, Gaussian filters and geometric filters.
Several semiautomatic segmentation methods in ultrasound images have been
proposed. These methods included pixel and region based segmentation, edge
based segmentation and hybrid techniques [2][3]. The success of these techniques
is moderate because they cannot obtain a precise shape of the region of interest
due to the special characteristics of ultrasound images or the requirement of a
priori knowledge of the shape.
Active contours technique, also called snakes, is recognized as one of the most
efficient tools for image segmentation [4]. This method consists in deforming an
initial contour of the object under a set of internal and external forces. Several
difficulties appear when applying this model to ultrasound images because of
ultrasound characteristic noise, the requirement of an initial outline of the nodule
boundary to start the algorithm and the bad adequacy of the external forces that
guide the snake.
In order to reduce the speckle noise which characterizes ultrasound images,
we have used the truncated median filter. In a few iterations, the image is much
more suitable for the application of a region growing algorithm. On the filtered
image, we apply the structure tensor to estimate the magnitude of the gradient.
Fig. 1. From top to bottom and from left to right: Original image, filtered image,
magnitude of the gradient using the structure tensor and initial segmentation
From a seed point introduced by the specialist, the selection grows until a certain
threshold for the magnitude of the gradient is reached (see Fig. 1).
Afterwards, the contour of the selected area is considered as the initial snake
for the evolution of the active contours, based on the following equation:
∂u(x, y) 5u(x, y)
= g(I(x, y)) |5u(x, y)| div( ) + λ∇g(I(x, y)) 5 u(x, y) (1)
∂t |5u(x, y)|
where u(x, y; 0) = u0 (x, y) is the initial snake contour, λ ≥ 0 is a constant

that guides the internal force of the snake and g(I(x, y)) represents the stop
function, the external force that attracts the snake u(x, y) to the real contour
of the object. It is an edge detector that must be selected according to the
characteristics of the image.
Fig. 2. From top to bottom and from left to right: Initial segmentation, final snake using
Gaussian filtering without parameter evolution, final snake using Gaussian filtering
with parameter evolution and final snake using geometric filters
A typical choice of g(I(x, y)) is:
1
g(I(x, y)) = q (2)
1 + α |∇Is (x, y)|2
where α ≥ 0 is a constant and Is (x, y) denotes the smoothed version of

the original image. We have used a multiscale implementation of the equation
proposed in [5] that allows adapting the external and internal forces for each
scale.
We have adapted snake adjustment from the classical model to the special
characteristics of ultrasonography, and we have compared the most representa-
tive methods for speckle reduction in ultrasound images. Initially, the best results
have been obtained using geometric filters. The non-linear geometric algorithm
which has been selected uses the Crimmins complementary hulling algorithm [6].
We have tested the proposed scheme on a set of breast ultrasound images,
comparing the results with the classical snake implementations (see Fig. 2). The
comparison of the results provided by the system and those supplied by the
specialists through manual delineation shows the accuracy and reliability of the
proposed technique, and emphasizes its usefulness for the further analysis of the
diagnostic criteria used in the classification of the nodules.
References
1. Stavros, A.T., Thickman, D., Rapp, C.L., Dennis, M.A., Parker, S.H., Sisney, G.A.:
Solid breast nodules: use of sonography to distinguish between benign and malignant
lesions, Radiology 196 (1995) 123-134
2. Drukker, K., Giger, M.L., Horsch, K., Kupinski, M.A. and Vyborny, C.J.: Comput-
erized lesion detection on breast ultrasound. Medical Physics 29:7 (2002) 1438-1446
3. Revell, J., Mirmehdi, M. and McNally, D.: Applied review of ultrasound image fea-
ture extraction methods. In A Houston and R Zwiggelaar, editors, The 6th Medical
Image Understanding and Analysis Conference, BMVA Press (2002) 173—176
4. Kass, M., Witkin, A. and Terzopoulos, D.: Snakes: Active Contour Models. In 1st
International Conference on Computer Vision (1987) 259-268
5. Caselles, V., Kimmel, R., Sapiro, G.: Geodesic Active Contours, International Jour-
nal of Computer Vision, 22:1 (1997) 61-79
6. Busse, L.J., Dietz, D.R., Glenn, W.M. jr: A non-linear algorithm for speckle reduc-
tion, IEEE Symposium on Ultrason. Ferroelec. and Freq. Contr., IEEE 91CH3079-1
(1991) 1105-1108
Parallel Progressive Multiple Sequence
Alignment
Erik Pitzer
University of Applied Sciences Hagenberg, 4232 Hagenberg, Austria,

Erik.Pitzer@fh-hagenberg.at
Sequence alignment is one of the most basic tools in biological sequence

analysis, used to compare sequences or infer their evolution. While alignments for
two sequences can readily be calculated, the problem of producing exact multiple
sequence alignments is almost incomputable. Therefore, several algorithms were
developed to accelerate the calculation by reducing the complexity from multiple
dimensions to a series of two dimensional problems. This approximation makes
the problem solvable in a reasonable amount of time. Still, the procedure is very
costly. As a consequence, we tried to accelerate the progressive multiple sequence
alignment by exploiting inherent parallelism. Although the algorithmic system
of progressive multiple sequence alignment appears monolithic at first glance, it
contains a multitude of concurrently executable paths. We found that all stages of
the algorithm are highly scalable and that an alignment problem of n sequences
can be distributed in a way to employ at least n computers or even n2 computers
during the preliminary stages of pairwise alignment and phylogenetic inference.
Sequence alignment provides a way to determine similarity between sequences.
To accomplish this, usually some form of dynamic programming is used. A ma-
trix, which contains the solutions to sub-problems, is calculated. From this ma-
trix the overall solution can be derived. This implies a complexity of O(n2 ). If
more sequences are involved, the matrix becomes multidimensional. This expo-
nential growth in complexity is, however, hardly maintainable.
To obtain at least an approximation, accuracy is sacrificed for speed. These
approximation algorithms use a technique called progressive multiple sequence
alignment, which is a consecutive application of pairwise sequence alignment al-
gorithms. By extending the algorithms to additionally consider evolutionary,
physical and chemical processes the obtained results are often more reason-
able than an exact solution could have been. A very popular implementation
‘Clustal W’ is described in [1] and [2].
As the progressive method has several stages that introduce gaps in the se-
quences, the right choice of order is crucial to obtain meaningful results. During
the progressive process, some sequences are aligned pairwise. The calculated re-
sults are then treated as if they were single sequences again. In the next step they
are combined into bigger alignment sub-groups, until all sequences have been in-
corporated into a final alignment. The correct order is obtained by constructing
a preliminary tree of evolution, a so-called guide tree, which is usually obtained
by means of the Neighbor-Joining method [3]. This method, in turn, operates on
a matrix of distances between all sequences. Thus, the progressive multiple se-
quence alignment actually consists of three stages. Each resulting stage depends
on the complete solution to the previous stage, so there is no chance of execut-
ing the stages in parallel. The stages themself, however, can be parallelized very
well.
In the first stage, the pairwise distances between the sequences are calcu-
lated. For every pair of sequences a pairwise sequence alignment has to be cal-
culated. This results in n2 pairwise alignment problems. These problems are
all completely independent of each other and can be easily distributed using the
master-worker pattern. Because of this simplicity, the class of this problem is also
called ‘embarassingly parallel’. We developed a framework that distributes tasks
in such a fashion. The greatest advantage over predetermined task scheduling is
its dynamic adaptation to computer and network speed that will, on average,
be optimal for any given network architecture. Additionally, the task scheduler
can conglomerate tasks into jobs that pertain to a specific alignment problem.
In this manner, the computer cluster can calculate several sequence alignments
in parallel. The scheduler also honors the total complexity of these jobs, so that
simpler jobs can ‘overtake’ bigger ones. In this manner, the average processing
and waiting time can be reduced.
The result of pairwise alignments stage is a distance matrix containing the
pairwise distances of all sequences. This is the input for the Neighbor-Joining
method which is used to build a preliminary tree of evolution. The Neighbor-
Joining process incurs a repeated re-evaluation of the resulting tree for every
possible pair of nearest neighbors. In this manner, the age of the sequences is also
considered. In fact, the Neighbor-Joining process is able to exactly reconstruct
any tree structure with only the distances of leaves as the input. As every step
includes the consideration of n2 pairs, for which all distances in the tree have to
be evalutated, it is again perfectly divisible into independent tasks. After every
search for the next neighbor the one with the best tree (the tree with the smallest
average distance between the leaves) is chosen. This is done by an intermediate
synchronization step, where all participants ‘publish’ their results and continue
all with the optimal neighbors. Therefore, every computer has all data locally
available at any time, and there is very little data that has to be transferred for
synchronization.
The last stage is actually the most interesting one. While the progressive
sequence alignment can only be parallelized well in the beginning, when there
are still many pairs of single sequences, the alignment itself of two sequences
(and its application to groups) is applied during all iterations of the progressive
alignment process. Therefore, its acceleration is much more crucial. The dynamic
programming approach does not exhibit as many independent tasks as the other
two stages. Every cell in the matrix depends on three other cells that represent
the sub-solution to smaller problems. Nevertheless, the first row and the first
column can be calculated immediately. Actually, after the first element has been
calculated, there are two other elements that can be calculated: the element to
its right and the element below. And after that, there are already three elements
that can be calculated. Hence, instead of sequentially moving through the matrix
rowwise or columnwise, a diagonal procession of execution can be built up that
will employ n/2 computers on average. To preserve time or space, whichever is
most important, the algorithm can be dynamically chosen to perform a sequence
alignment in the style of Needleman-Wunsch [4] or Myers-Miller [5].
The parallelization of progressive multiple sequence alignment bears great po-
tential in exploiting the computing power of many machines. The availability of
solutions to bigger and bigger sequence alignment problems makes this algorithm
immediately applicable to practical problems. Moreover, the implementation of
the discovered concurrency methods could be applied to the parallelization of
graph and matrix algorithms in general.
References
1. Higgins, D.G., Sharp, P.M.: CLUSTAL: a package for performing multiple sequence
alignment on a microcomputer. Gene 73 (1988) 237–244
2. Thompson, J.D., Higgins, D.G., Gibson, T.J.: CLUSTAL W: improving the sensitiv-
ity of progressive multiple sequence alignment through sequence weighting, position-
specific gap penalties and weight matrix choice. Nucleic Acids Research 22 (1994)
4673–4680
3. Saitou, N., Nei, M.: The Neighbor-joining Method: A New Method for Recon-
structing Phylogenetic Trees. Journal of Molecular Biology and Evolution 4 (1987)
406–425
4. Needleman, S.B., Wunsch, C.D.: A General Method Applicable to the Search for
Similarities in the Amino Acid Sequence of Two Proteins. Journal of Molecular
Biology 48 (1970) 443–453
5. Myers, E.W., Miller, W.: Optimal alignments in linear space. Computer Applica-
tions in the Biosciences 4 (1988) 11–17
Computer Aided Simulation of Energy Limits
and Development of Bio-Thermodynamic Systems
Stanislaw Sieniutycz
Faculty of Chemical Engineering, Warsaw University of Technology,

00-645 Warsaw, 1 Warynskiego Street, Poland
sieniutycz@ichip.pw.edu.pl
Extended Abstract
In this research we focus on the theory and computer aided modeling of energy limits
in bio-thermodynamic systems with living organisms as the effect of extremum
properties of entropy and its associates. We also consider the role of physical laws in
establishing optimal forms and sizes of the organisms, and stress the universality of
these laws for animate and inanimate systems. For biological processes with variable
number of states an extremum principle provides quantitative picture of their
development. The results show a discrete gradient dynamics is governed by the
entropy potential. For an accepted complexity, some evolutions may progress in a
relatively undisturbed manner, whereas others may terminate in a rapid way due to
inherent instabilities, the features of variational principles for shortest paths and
transversality conditions. Numerical aspects of energy and size limits are described.
Recently, several approaches involving systems theory, computer science, optimal
control theory and thermodynamics have been applied, aimed at evaluation and
simulation of energy and size limits for bio-thermodynamic systems [1]. Static limits,
which refer to steady situations, were evaluated earlier; more complex limits
associated with dynamic processes were found later [1, 2]. Static limits may be
defined as maximum mechanical work produced or minimum work consumed by a
complex system or an organism at a steady state. Dynamic limits may be referred to
the same physical or biophysical quantities as static, yet they may include other terms,
e.g. those characterizing rates of the system’s evolution. Often, dynamic limits are
defined through extremum values of certain integrals, e.g. those describing action,
entropy production or availability dissipation. Usually dynamic limits are associated
with the system’s (organism’s) development, including phases of growth and aging.
Typical constraints at which the limits are operative are energy and mass balances.
Both limits are numbers referred either to steady-state points or to the whole evolution
of the system. Yet, static limits are functions of a steady-state point, whereas dynamic
limits are functionals of the strategy the system develops. In fact, the dynamic limits
are most often determined as extremum values of respective functionals.
In the course of its development, the system (organism) can change not only its
mass-energy outputs (inputs), but also the size, structure and shape. Size limits are
defined in terms of the extremum size that it can achieve in the course of its
individual development or during its evolution characterizing the progress through
generations. Growing systems (those that are capable of increasing their size and/or
number of states) can exhibit critical behavior when their size increases beyond a
certain value. In effect, some developments may progress in a relatively undisturbed
manner, whereas others may terminate in a rapid way due to instabilities. For majority
of nonlinear models their solutions may only be numerical; so we describe also basic
numerical aspects of calculations associated with the computer modeling of the limits.
The structure of the research is as follows. In its basic part energy and size limits
are treated. In particular, limits for steady power yield or consumption are considered
and a problem of dynamical energy limits is formulated. This part terminates with the
scheme of multistage production of mechanical energy (work) in a sequence of
animate or inanimate energy generators. Next we refer to a group of works [1, 3] that
analyze the evolution of living organisms treated as multistage systems by the
complexity criterion based on a potential, this potential being the entropy itself or an
entropy-based complexity. Classical thermodynamic quantities do not appear, yet the
model yields an extremum principle that, as in thermodynamics, implies an extremum
of a potential. Dual forms of this principle are also considered. Next we develop a
discrete, nonlinear generalization of dynamics in metric spaces that may be
curvilinear. Numerical approaches applying the dynamic optimization methods and
leading to unsteady-state limits are then described and, in particular, sequential
systems with growing number of stages are treated by methods of dynamic
programming and maximum principle. Suitable dynamic programming algorithms
(Bellman’s recurrence equations) are derived and computer-aided simulations of their
solutions are discussed. In the final part we conclude that many features of living
organisms can be predicted by describing their complex evolutions in terms of
variational principles for shortest paths along with suitable transversality conditions.
Energy, size and shape limits. Each organism can be regarded as a complex system
with very many elementary objects, links and couplings. Some inputs serve to supply
the consumed energy and substances to the organism (system) whereas some outputs
(e.g. muscles) are able to deliver produced mechanical energy out of the organism. In
fact, there is an analogy between an energy system of inanimate Nature and a living
organism. There are also at least two essential differences between the living system
and a machine: the first one is the much larger complexity of living system than a
machine and the second is the presence of huge number of information streams out
and within the living system. Yet, the essential property of living systems, the
capability to work, is typical of work generators. By applying physical principles we
are nowadays able to evaluate energy and dimension limits for living organisms and
describe behavior of many organs. While living systems have developed various
strategies to manipulate their organization in order to satisfy the principles of both
physics and biology, it is the physical laws that set limits to their size, functioning and
rate of development. For example, the physical law of thermal conduction sets the
size of warm-blooded aquatic animals that require a minimum diameter in order to
survive in oceans. Land animals that are able to run have body masses ranging from
little mice of M= 10-2 kg to big ungulates weighing about half ton (0.5x103 kg).
Aquatic mammals span the range from M= 101 kg baby seals to 105 kg blue whales.
Flying birds size from 10-2 kg humming birds to 15 kg mute swans [4]. These results
are adduced following Ahlborn [4] who has shown that the size limitations are
consequence of thermodynamics which sets the energy budget and power production
of animals. The power of all animals in terms of their body mass can be described by
a unique metabolic law. Species in ecosystems funnel energy into their own
production or reproduction and influence catalytic processes in the ecosystem.
A number of recent developments in thermodynamic optimization focus on the
generation of optimal geometric form (shape, structure, topology) and shape limits in
animate and inanimate flow systems, where the flow configurations are free to vary.
The principle that generates geometric form searches for best global performance (e.g.
minimum irreversibility) subject to global finiteness constraints (volume, weight,
time). The resulting structures obtained in this way are named constructal designs [5].
The same objective and constraints principle applies for the optimally shaped flow
paths that occur in natural systems (animate and inanimate) which is Bejan’s
constructal theory [5]. Examples are obtained from engineering: the distribution of
heat transfer area in power plants, optimal sizing and shaping of flow channels and
fins, optimized ratios of heat exchanger structures, aerodynamic and hydrodynamic
shapes, tree-shaped assemblies of convective fins, tree-shaped networks for fluid
flows, optimal configurations for streams with bifurcation, insulated pipe networks
distributing hot water and exergy over a fixed area, and distribution networks for
entities transferred in societies (goods, currency, information). The principle-based
generation of flow geometry unites optimization developments known in mechanical
engineering with those in civil engineering and social organization. These results,
obtained in A. Bejan‘s research group [5], generalize methods of thermodynamics, as
they show how their principles can mimic the development of optimal configurations
in engineering and social organization. The organization is understood in a broader
sense, in particular as the one that refers to populations of living organisms (e.g. those
in ecosystems). The results show that flow systems with geometric structure exhibit at
least two flow regimes, one with high resistivity (slow, diffusion, walking) that fills
the volumes of smallest finite scale, and one or more with low resistivity (e.g., fast,
streams, channels, streets). The optimal balancing of the regions with different flow
regimes means that the material and channels must be distributed in certain ways.
Better global performance is achieved when the distribution is relatively uniform, in
spite of the large differences between the high- and the low-resistively domains [5].
Development and evolution of species. It is the difference between the maximum
entropy and the actual (observed) entropy that quantifies the organization in a
complex system. This property refers to the various forms of organization: with
respect to characters over species, cells over all cell types, organisms over genotypes.
Moreover, it is the entropy-based complexity, rather than fitness or organization, that
determines direction of evolutionary processes. Spontaneous increase of complexity
is regarded as a basic feature of evolution. For biological systems a reasonable
measure of complexity is the different number of components they contain, so as to
be consistent with well-known Williston's law which predicts that very similar
components will either merge or specialize. Staying on the ground of the idea of
increasing complexity, the system's organization acts as a force which prevents loss
of components and allows complexity increase by the integration of new ones. When
speaking about the complexity and related entropy one issue is particularly important:
a nonequililibrium entropy has to be necessarily applied because the difference
between the maximum entropy and the actual (observed) entropy governs the
organization in a complex system. Here we consider the idea of the extremum
development of complexity or complexity-related criterion of evolution as applied to
some recent information-theoretic models for multistage living systems characterized
by sequentially evolving states. For these models the evolution of living organisms is
a multistage effect governed by the complexity criterion Γ based on information-
theoretic entropy. An original optimization problem of maximum change of entropy
or entropy-related complexity Γ between a point (for a system with n states) and the
surface of constant distance from the point can be formulated and solved by
variational calculus. For practical purposes also a dual of this problem is considered
which is the problem of minimum distance between a point (for a system with n
states) and the entropy manifold of the system with n+1 states. In this formulation, an
s-dimensional evolution can be obtained via minimization of the length functional
where the independent probabilities pi are constrained to reside on constant-
complexity manifold satisfying the constraint dΓ=0. It may then be shown that both
the original problem of maximum complexity and its dual problem of the minimum
distance yield a discrete evolution dynamics satisfying the Onsager-like form
∂Γn+1( p0,. ps) ∂Γn+1( p0,. ps) ∂Γn+1( p0,. ps)
ps (n +1) − ps (n) = Ls1 + Ls2 ...+ Lss , (1)
∂ p0 ∂ p1 ∂ ps
where Lik = gik∆τ. The symmetry matrix Lik is the well-established result of
irreversible thermodynamics. Therefore the evolution process can be imbedded quite
naturally into a large family of thermodynamic processes.
Conclusions. We defined energy, size and shape limits and extremum properties of
complex systems (living organisms). We described numerical aspects of calculations
associated with the computer aided modeling of these limits. We have shown that the
development of living organisms can be described in terms of variational extremum
principles along with suitable transversality conditions. We found gradient dynamics
that governs the evolution problems, the result that is consistent with the entropy
principle of extremality as the driving factor in the discrete dynamics of complex and
living systems, postulated recently [3]. We conclude that the extremum principle
should be integrated with the theory of behavior of biochemical systems.
Acknowledgments. The author acknowledges the financial support from the KBN,
grant T09C 024 26: Non-equilibrium thermodynamics and optimization of chemical
reactions in physical and biological systems.
References
1. Berry, R.S., Kazakov V.A., Sieniutycz S., Szwast Z., Tsirlin A. M.: Thermodynamic
Optimization of Finite Time Processes, Wiley, Chichester (2000).
2. Jorgensen, S.E., (ed.): Thermodynamics and Ecological Modelling, Lewis Publishers of
CRC Press, Boca Raton (2001).
3. Gontar, V.: Entropy principle of extremality as a driving force in the discrete dynamics of
complex systems, Chaos, Solitons and Fractals 11 (2000) 231-236.
4. Ahlborn B.K., Thermodynamic limits of body dimension of warm-blooded animals, J. of
Non-Equilib. Thermodyn. 40 (1999) 407-504.
5. Bejan, A., Shape and Structure from Engineering to Nature, Cambridge University Press,
Cambridge (2000).
Soft Computing and Geometrical Control for
Computer Aided Driving
Javier Alonso Ruiz, Teresa de Pedro, Carlos González y Ricardo García
Industrial Automation Institute, CSIC, La Poveda, Arganda del Rey.

28500 Madrid, Spain
{jalonso, ricardo, tere, gonzalez}@iai.csic.es
Abstract. After having completed the design of control systems for real
autonomous cars in urban environment using straight lines as reference, we are
now intending to build a fuzzy control system based in clothoids [2], the curve
used in roads and train ways. This work proposes a method based on soft
computing and upgraded via genetic algorithms. Both sets of simulations are
shown, and compared.
Introduction
One of the big efforts in a control system is to reduce the force applied on the
actuators. In autonomous vehicles the amount of electrical required is an important
issue. Our goal is to reduce the electrical expense in the direction actuators. Moreover
the reduction of excessive steering wheel movements increases the comfort feeling.
We have proposed and tested a control method for vehicle navigation in curved
sections. We have used a curvature control method based on fuzzy logic with genetic
algorithms optimization over a geometric trajectory design. We have obtained
impressive results on simulation and now we are testing these results on real
autonomous vehicles.
Nowadays most roads are designed trying to maintain a progressive and smooth
trajectory in its curved sections. This designing method is explained in “Carreteros”
web page [4]. Each curve is divided into three parts, in the first one the curvature of
the road increases, in the second one the curvature is maintained constant and in the
last part the curvature is reduced to cero. Based on this idea we design a trajectory for
each curve. Following this trajectory the vehicle smoothly changes its curvature from
cero to a given number, and then from this number to cero.
To be more precise in [3] this differential equation system is shown relating the state
and control variables and its integration equations for admissible paths:
(x ′ y ′ θ ′ κ ′) = v ⋅ (cos θ sin θ κ 0 ) + σ ⋅ (0 0 0 1) (1)
where (x y θ κ) stand for attitude and curvature, and σ is σ= κ’= Φ/cos2Φ. (Φ is the
wheel angle).
Our proposal is to express position of the car depending on its curvature,
f(x,y,κ,v,t) related to steering angle and control the car using its usual inputs,
acceleration and curvature (v’, κ=1/r).
In order to obtain a solution as smother as possible we have used a fuzzy logic

control system [1]. This control system must maintain the curvature goal that depends
on the point of the curve where the car is. It will work on the curvature error, its first
instant derivation and the error accumulated on that curve, just to minimize the
curvature error. So we use a fuzzy PID fuzzy control.
The membership function parameters are improved with a genetic algorithm fitness
function that combines the accumulated effort and the maximum error made on the
circuit curves. We have obtained different optimized set values depending on the
nature of the circuit curves (motorways with small curvature or race circuits).
The simulator uses detailed maps of competition circuits, mainly Xerez, with its
real proportions. And it takes care of our cars, CITROËN, mechanical characteristics.
We have carried out several simulations in this circuit, and the results show a drastic
effort reduction after the genetic algorithm optimization.
Left fig. Differences between two consecutive steering wheel actuator goals. The blue one is
the steering wheel actuation without optimization and the pink one is after application of
genetic algorithm. Right fig. Drawing a course on Xerez circuit map
References
1. García R and de Pedro T, “First Application of the ORBEXoprocessor : Control of

Unmanned Vehicles” Mathware and Soft Computing, n. 7, vol 2-3, pp 265-273, 2000
2. Jiménez Shaw J, “Resolución numérica completa de la Ecuación de la Clotoide”, web
page and Clothos constants calculation program. http://javier.jimenezshaw.com/
3. Laugier C and Fraichar T, “Decisional Architectures for Motion Autonomy” chapter 11in
Intelligent Vehicle Technologies, editors: Vlacit, ParenT, Harashima, SAE International,
ISBN 0 7680 0780 I
4. Web page www.carreteros.org, Norm 3.1-IC “Trazado”
A Monocular Solution to Vision-based ACC in Road
Vehicles
Jesús Nuevo1, Miguel Ángel Sotelo1, Manuel Ocaña1, Luis Miguel Bergasa1
1Department of Electronics, University of Alcalá, Alcalá, Madrid, Spain
{jnuevo, michael, mocana, bergasa}@depeca.uah.es
Abstract. This paper describes a monocular vision-based Adaptive Cruise Con-

trol (ACC) System in the framework of Intelligent Transportation Systems
(ITS) technologies. The challenge is to use a single camera as input, in order to
achieve a low cost final system that meets the requirements needed to under-
take serial production.
1 Introduction
A monocular imaging device (a single FireWire digital camera) is deployed to pro-

vide “indirect range” measurements using the laws of perspective. Some previous
developments use available sensing methods such as radar [1], stereo vision [2], or a
combination of both [3]. Only a few works deal with the problem of monocular vehi-
cle detection using symmetry and color features [4], or pattern recognition techniques
[5]. In the current work, the searching space is reduced in an intelligent manner in
order to increase the performance of the detection module. Accordingly, road lane
markings are detected and used as the guidelines that drive the vehicle searching
process. The area contained by the limits of the lanes is scanned in order to find vehi-
cle candidates that are passed on to the vehicle recognition module. This helps reduce
the rate of false positive detections. In case that no lane markings are detected, a basic
area of interest is used instead covering the front part ahead of the ego-vehicle. The
description of the lane marking and vehicle detection systems is provided below,
together with some graphical results.
2 System Description
Lane Tracking
The system is divided in three modular subsystems with specific functions. The
first subsystem is responsible for lane detection and tracking, as well as lane crossing
monitoring. Images obtained from the camera are processed and clothoid curves are
fitted to the detected markings. The algorithm scans up to 25 lines in the area of in-
terest, from 2 meters in front of the camera position to below the horizon. The devel-
oped algorithm implements a non-uniform spacing search that reduces certain unsta-
bilities in the fitted curve. The final state vector is composed of 6 variables [7] for
each line on the road: coh, c1h, cov, c1v, xo, Ρo, where coh and c1h represent the clothoid
horizontal curvature parameters, cov and c1v stand for the clothoid vertical curvature
parameters, while xo and Ρo are the lateral error and orientation error, respectively,
with regard to the centre of the lane. The clothoid curves are then estimated based on
lane marking measurements using a Kalman filter for each line. These lines conform
the area of interest. Figure 1 depicts a sequence of images in which the result of the
lane tracking algorithm is overprinted on the road images.
Fig. 1. Lane tracking example in a sequence of images. The green lines represent the estimated
lines of the road. The example also depicts the error between the left wheel of the car the the
left lane (left), the error between the right wheel of the car and the right lane (right), the radious
of curvature of the road estimated at a lookahead distance of 50m (R), and the maximum rec-
ommended velocity to bend the curve (V) according to the radious of curvature.
Car Detection and Recognition
The area of the image contained between the lines that represent the lane is
scanned in order to look for candidate vehicles along the lane as depicted in figure 2.
Car detection is performed in the areas limited by lane markings. Based on character-
ics of shape, size and position, among others, candidates are selected. Each one of
the candidates is then divided into five different subareas, that are then processed
independently. For each subarea of the candidate, a support vector machine (SVM)
[6] has been trained, using the tool ‘TsetBuilder’ we have developed. The areas are
validated independently, and another SVM determines whether the outputs of the first
SVMs represent those of a valid cadidate. Candidates are tracked using a Kalman
filter. Figure 3 shows the result of the detection and tracking algoritm.
Fig. 2. Sequential vehicle candidates searching along the detected lane
Fig. 3. Vehicle tracking example in a sequence of images.
The distance between the ego-vehicle and the preceding vehicle along the lane be-
comes the input to the Adaptive Cruise Control System.
Acknowledgements
This work has been funded by Research Projects CICYT DPI2002-04064-05-04

(Ministerio de Educación y Ciencia) and FOM2002-002 (Ministerio de Fomento,
Spain).
References
1. G. R. Widman, W. A. Bauson, and S. W. Alland, “Development of collision avoidance

systems at Delphi Automotive Systems”. In Proc. Int. Conf. Intelligent Vehicles, pp. 353-
358, 1998.
2. T. Williamson and C. Thorpe, “Detection of small obstacles at long range using multibase-
line stereo”. In Proc. Int. Conf. Intelligent Vehicles, pp. 311-316, 1998.
3. R. Labayrade, C. Royere, D. Gruyer, and D. Aubert, “Cooperative fusion for multi-obstacles
detection with use of stereovision and laser scanner”. In Proc. Int. Conf. On Advanced
Robotics, pp. 1538-1543, 2003.
4. A. Broggi, M. Bertozzi, A. Fascioli, C. Guarino Lo Bianco, and A. Piazzi, “The Argo
autonomous vehicle’s vision and control systems”. International Journal of Intelligent Con-
trol and Systems. Vol. 3, No. 4, 409-441, 2000.
5. G. P. Stein, O. Mano, and A. Shashua, “Vision-based ACC with a single camera: bounds on
range and range rate accuracy”. In Proc. Int. Conf. Intelligent Vehicles, 2002.
6. Christopher J.C. Burges, “A Tutorial on Support Vector Machines for Pattern Recognition”.
Data Mining and Knowledge Discovery, 2,121-167 (1998). Kluwer Academic Publishers.1.
7. Dickmanns E. D. and Mysliwetz. B. D. “Recursive 3-D Road and Relative Ego-State Rec-
ognition”. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 14, No. 2,
February 1992.
Automatic Parallel Parking for Autonomous
Vehicles
Javier de Lope and Darı́o Maravall
Department of Artificial Intelligence

Faculty of Computer Science
Universidad Politécnica de Madrid
Campus de Montegancedo, 28660 Madrid, Spain
{jdlope,dmaravall}@dia.fi.upm.es
Extended Abstract
Optimization is one of the most common and pervasive issues in real-world engi-
neering and economic systems and it is at the heart of any decision-making task
in which a choice must be made between several alternatives to achieve multi-
ple, sometimes conflicting, objectives. The objectives are generally formalized as
analytical functions or performance indices.
The technical literature on optimization methods is really extensive, as this
fundamental subject has received a tremendous amount of attention since it
came of age about 50 years ago. Single-objective optimization is by far the most
researched problem in this field, although many real-life situations are multi-
objective optimization problems per se. A multi-objective optimization problem
is very often converted into a single-objective optimization case by integrating
the multiple performance indices into a single one [1].
The standard solution for truly multi-objective optimization problems is to
find the so-called Pareto-optimal front. The Pareto front is formed by the solu-
tions in which any change in any of the decision variables aimed at improving a
particular performance index will produce a deterioration in some of the other
performance indices. Due to the inherent difficulties in calculating the analytical
Pareto optimal surfaces for many real-world systems, evolutionary methods have
lately been applied to solve multi-objective optimization problems [2, 3].
In most of the multi-objective optimization problems that have been tackled
by evolutionary techniques so far, the performance indices to be optimized:
Ji (x) = Ji (x1 , x2 , . . . , xn )
(1)
x ∈ Rn ; i = 1, 2, . . . N
are not time-varying and the decision varibles x1 , x2 , . . . xn have static con-
straints:
gi (x) ≥ 0 ; i = 1, 2, . . . p
(2)
hi (x) = 0 ; i = 1, 2, . . . q
Only very recently have a number of papers dealing with the application
of evolutionary computation to solve multi-objective optimization cases been
published, in which either the performance indices Ji (x), the decision variables
x or both are time dependent and dynamic [4, 5]. Most of this research work
refers to variations on static optimization problems, in which some well-known
static benchmark optimization functions are exposed to time-varying dynamics.
Noise in the measurement of the performance indices or in the decision variables
is sometimes also added.
In this paper, we present an engineering problem (automatic car parking)
that ultimately leads to the formalization of an active multi-objective dynamic
optimization problem as a cooperative game. However, instead of using evolu-
tionary controllers to search for an optimal, or at least efficient, set of if-then
reasoning rules, we propose a much simpler method based on what we call a
biomimetic approach to sensory-motor coordination. Sensory-motor coordina-
tion is a central issue in the design of most real-world engineering systems, as
the designer somehow has to deal with the coordination of perceptual informa-
tion and physical actions. We have already applied this biomimetic approach
to solve specific manipulation and locomotion problems in articulated mecha-
nisms [6, 7] and wheeled vehicles [8, 9]. In this paper we illustrate our approach
to sensory-motor coordination with the problem of automatic parallel parking
for autonomous vehicles.
References
1. Deb, K. (2001) Multi-Objective Optimization Using Evolutionary Algorithms. Wi-
ley, New York
2. Fonseca, C.M., Fleming, P.J. (1993) Genetic algorithms for multiobjective opti-
mization: Formulation, discussion and generalization. Proc. of the 5th Int. Conf. on
Genetic Algorithms, 416–423
3. Coello, C.A. (2003) Special Issue on Evolutionary Multiobjective Optimization.
IEEE Trans. on Evolutionary Computation, 7:2, 97–99
4. Branke, J. (2002) Evolutionary Optimization in Dynamic Environments, Kluwer,
Boston
5. Jin, Y., Sendhoff, B. (2003) Connectedness, regularity and the success of local search
in evolutionary multi-objective optimization. Proc. IEEE Congress on Evolutionary
Computation (CEC-2003), 1910–1917
6. Maravall, D., De Lope, J. (2003) A bio-inspired robotic mechanism for au-
tonomous locomotion in unconventional environments. In C. Zhou, D. Maravall, D.
Ruan (Eds.) Autonomous Robotic Systems: Soft Computing and Hard Computing
Methodologies and Applications. Physica-Verlag, Springer, Heidelberg, 263–292
7. De Lope, J., Maravall, D. (2004) A biomimetic approach to the stability of biped
robots. Climbing and Walking Robots. Springer-Verlag, Heidelberg
8. De Lope, J., Maravall, D. (2003) Integration of reactive utilitarian navigation and
topological modeling. In C. Zhou, D. Maravall, D. Ruan (Eds.) Autonomous Robotic
Systems: Soft Computing and Hard Computing Methodologies and Applications.
Physica-Verlag, Springer, Heidelberg, 103–139
9. Maravall, D., De Lope, J., Patricio, M.A. (2004) Competitive Goal Coordination
in Automatic Parking. Proc. 1st of the European Workshop on Evolutionary Algo-
rithms in Stochastic and Dynamic Environments (EvoSTOC-2004)
Intelligent Automation of Electric Power Steering
Systems
J. E. Naranjo, C. González, R. García, T. de Pedro
Instituto de Automática Industrial

Ctra. Campo Real Km. 0,200, La Poveda, Arganda del Rey, Madrid 28500
Tlf. +34 918711900, Fax +34 918717050
{jnaranjo, gonzalez, ricardo, tere}@iai.csic.es
Abstract. The automatic control of the steering wheel of a vehicle is presently

one of the most important challenges in the Intelligent Transportation Systems
field. In this paper we present an automatic steering control system for mass-
produced electric steering wheel equipped vehicles, based on fuzzy logic, that
allows a behavior very similar to the human are. In literature, we can find a lot
of theoretic proposals and some simulations, but only a few working teams of-
fer real solutions for this task. One of this is the work developed by the
AUTOPIA Program in which some vehicles have been automated allowing per-
forming some maneuvers mimicking the human reactions. In the present work,
we use the pre-installed electrically assisted steering wheel of a vehicle Citroën
C3 Pluriel in order to act over the attitude of the car. This actuator is controlled
from a onboard computer in which is executing a fuzzy logic based autonomous
steering system. The internal computers of the vehicle generate the input infor-
mation, that is read through a CAN bus, and a high precision GPS. With this
equipment we have performed some experiments in a private test circuit show-
ing a human-like behavior in all the maneuvers.
1 Introduction
The development of Intelligent Transportation Systems (ITS) provides an opportunity

to apply advanced technology to systems and methods of transport for efficient, com-
fortable and safer highways, railways, inland waterways, airports, ports and linkages
between these different types of transport. We center our work in the area of road
transport, more specifically in the intelligent vehicles sub area, which includes the
topic of autonomous vehicles. This topic refers to vehicles that equip the necessary
instrumentation and intelligence for supplying the required service into the own vehi-
cle. This is, an autonomous car must self-control any or all of its functions without
external intervention. In this paper we present the automatic control of steering wheel
of a car. This way, in the Autopia program promotes its activities in this field, focus-
ing basically in autonomous driving, mainly based on fuzzy logic controllers. Steering
wheel [1], throttle [2] and brake pedal [3] automation have been carried out, working
with Citroën Berlingo vans and experimenting in private circuits. There are some
examples about automatic control of the steering wheel, as a step for achieving the
automatic driving. In the “Millemiglia in Automatica Tour” [4], a car was equipped
with a DC motor attached to the steering wheel through a pulley so it could be moved
depending of the commands of an onboard computer, in which resides an analytical
control system, that receives the sensorial input by artificial vision. In the Autopia
Program [5], the testbed vehicles vans have been automated in order to perform auto-
matic driving. This way, the steering wheel has also been modified so a DC motor
through gears could move it, and it is controlled from an industrial computer. In this
case, the sensorial input is provided from a Global Navigation Satellite System. The
necessity of attaching a motor to the steering wheel is because the assistance of the
conventional cars is powered by a hydraulic system. The newest vehicles equip Elec-
tric Power Steering (EPS), this is the steering assistance is made with an electric motor
that acts directly to the rack bar through a pinion. The motor torque depends on the
effort required by the driver. Some developments have been carried out in this way,
for example, the automatic parking system by Toyota [6]. The aim of this paper is to
present the control system for EPS developed for automatic driving and based on
fuzzy logic that has been installed in a Citroën C3 Pluriel vehicle (Fig. 1) in which we
have performed real experiments in a private circuit, obtaining human like behaviors.
Fig. 1. Citroën C3 Pluriel testbed vehicle.
References
1. R. García et al., “Frontal and Lateral Control for Unmanned Vehicles in Urban
Tracks”, IEEE Intelligent Vehicles Symposium, Versailles, France, 2002.
2. JE. Naranjo et. al., “Adaptive Fuzzy Control for Inter.-Vehicle Gap Keeping”,
IEEE Trans. on ITS,Volume 4: No. 3, September 2003, pp. 132-142.
3. J.E. Naranjo, et al., “A Throttle & Brake Fuzzy Controller: Towards the Automatic
Car”, LNCS 2809, Springer-Verlag, Julio 2003, pp 291-301.
4. A: Broggi et al., “The ARGO Autonomous Vehicles Vision and Control Systems”,
International Journal of ICS, Vol. 3, No. 4, pp. 409-441, 1999.
5. J.E. Naranjo et al., “Overtaking Maneuver Experiments with Autonomous Vehi-
cles”, Proc. 11th ICAR 2003, Coimbra, Portugal, pp. 1699-1703, June 2003.
6. T. Endo et al., “Development of reverse parking assist with automatic steering”,
ITS 2003 Congress Proceedings, Madrid, November 2003.
Development of an Artificial-Vision Controlled Traffic-
Light Prototype
Licesio J. Rodriguez-Aragón, Ángel Serrano, Cristina Conde, Raquel Montes Diez

and Enrique Cabello
Universidad Rey Juan Carlos (ESCET), C/Tulipán s/n,

28933 Móstoles, Spain
ljrodriguez@escet.urjc.es
http://frav.escet.urjc.es
Abstract.The design, development, construction and testing of an Artificial-

Vision controlled Traffic-Light prototype has been carried out to rule and regu-
late intersections. Methods, algorithms and automatons have been build up with
that purpose to provide the analysis of images and decisions making at real
time. The aim has been the development of an intelligent traffic-light capable
of capturing the presence or absence of vehicles, pedestrians and their particu-
lar situations defined by their trajectories. Besides the above mentioned proper-
ties we have to point out the adaptation to the precise characteristics of each
crossing, as its geometry, the required equipment, etc. The project has been su-
pervised by RACE, world wide known as experts in road safety awareness, en-
dowing the prototype with reliability and trust.
1 System Description
The system consists of two cameras situated at only one signal post placed at an inter-
section. One of them focuses on the pedestrian crossing while the other focuses on
the vehicles arriving grid, as seen in Figure 1. The prototype has been developed as to
be carried out by a conventional PC. A TWIDO programmable automaton, from the
Schneider Electric firm, is in charge of the traffic light control.
Fig. 1. General scheme of the prototype. Two cameras: one of them focussing on the pedestri-
ans and the other one on the vehicles. Images are then sent to the computer and as result a final
working mode is sent to the automaton responsible for the traffic light switch.
Two different types of algorithms have been developed. Low level algorithms are
in charge of obtaining the moving objects and assigning them into connected compo-
nents. High level algorithms are in charge of providing the system with crucial infor-
mation like position and trajectory, to make the required decision.
To evaluate and track the moving objects a background subtraction is performed

[1, 2]. The following step will be to delimit those objects. The calculation of the con-
vex hull of these objects has been finally established through a sequential labelling
while a matricial labelling algorithm was put aside due to the high computational cost
and the little improvement reached [3, 4].
Kalman filter is used for tracking moving objects, both vehicles and pedestrians.
An extended Kalman filter has been implemented [2, 5]. The algorithm adapts its
model at each step to improve the movement estimation of the component. Position,
velocity and values of the acceleration at different steps are used to predict the future
object location. Gathering the information provided by the trajectories and different
crucial zones established such as the crossing area and the vehicles grid, a decision
can be made on the operating mode to be sent to the automaton. These crucial zones
are fixed by initial parameters that have to be set in accordance with the particular
geometry of the crossing area.
2 RESULTS AND CONCLUSIONS
System works well under real conditions, without influencing the presence of
cameras in the behaviour of pedestrians and vehicles. After the multiple tests done
under RACE supervision, we can fairly state that the artificial techniques used in this
prototype are reliable and dependable to be introduced in our current daily life in
roads and pedestrians crossings. The enhancement of the methods and algorithms
used, allowing the decisions to be made in real time, make the system mature and
trustable enough.
Behaviour of vehicles in the grid can be analysed, and an option to control density
of traffic in different ways has been considered.
References
1 N. Papanikolopoulos, “Pedestrian Control at Intersections”, Intelligent transportation Systems Institute,

University of Minnesota, 2000.
2 N. Obolensky, D. Erdogmus, J. C. Principe, “An Time-Varing Kalman Filter to Moving Target Track-
ing”, Proc. of CONTROLO’02, pp 418-422. 2002.
3 R. Jain, R. Kasturi, B. G. Schunck, Machine Vision, McGraw-Hill, 1995.
4 R. C. Gonzalez, R. E. Woods, Digital Image Processing, Addison-Wesley, 1993.
5 P. J. Hargrave, “A tutorial Introduction to Kalman Filtering”, IEE Colloquium on Kalman Filters:
Introduction, Applications and Future Developments (Digest No.27), 1989.
Permanency Effect in Scene Depth Analysis
Miguel A. Fernández1, José M. López-Valles2, Antonio Fernández-Caballero1,

María T. López1, José Mira3 and Ana E. Delgado3
1Departamento de Informática, Escuela Politécnica Superior de Albacete
Universidad de Castilla-La Mancha, 02071 – Albacete, Spain
{miki, caballer, mlopez}info-ab.uclm.es
2 Departamento de Ingeniería de Telecomunicación, E.U. Politécnica de Cuenca
Universidad de Castilla-La Mancha, 16071 – Cuenca, Spain

josemaria.lopez@uclm.es
3 Departamento de Inteligencia Artificial, E.T.S.I. Informática,
Universidad Nacional de Educación a Distancia, 28040 - Madrid, Spain

{jmira, adelgado}@dia.uned.es
Extended Abstract
There are several strategies of how to retrieve depth information from a sequence of
images, like depth from motion, depth from shading and depth from stereopsis.
Brown et al. [1] describe a good approximation to all of them in their survey article.
Koenderink and van Doorn [2] expressed the necessary theory in their initial works
related to disparity restriction, and Wildes [3] implemented some of their ideas [4].
More recently, disparity in stereoscopy continues showing its great interest (e.g., [5]).
In this paper, as a novelty, we have chosen not to use direct information from the
image, but rather the one derived from motion analysis. The system proposed uses as
input the motion information of the objects present in the stereo scene, and uses this
information to perform a depth analysis of the scene.
The input to our system is a pair of stereo image sequences. The sequences have
been acquired by means of two cameras arranged in a parallel configuration. The
central idea behind this approach is to transpose the spatially defined problem of
disparity estimation into the temporal domain and to compute the disparity simultane-
ously with the incoming data flow. This can be achieved realizing that in a well-
calibrated fronto-parallel camera arrangement the epipolar lines are horizontal and
thereby identical to the camera scan-lines. Our team has already tested the motion
analysis algorithm used in this work in monocular video sequences ([6]-[8]).
In this case, motion analysis performs separately on both stereovision sequences in
two phases. The first analysis phase is based in grouping neighboring pixels that have
similar grey levels in closed and connected regions in an image frame. The method
used is segmentation in grey level bands. This method consists in reducing the resolu-
tion of illumination levels of the image, obtaining this way a lower number of image
regions, which potentially belong to a single object in motion. The second phase has
to detect possible motions of the segmented regions through the variation of the grey
level band of the pixels.
After motion detection, we now introduce a representation that may help to estab-
lish further correspondences between different motion information. This representa-
tion finds its basis in the permanency effect. This effect considers the jumps of pixels
between bands, and it consists of a matrix of charge accumulators. The matrix is
composed of as many units in horizontal and vertical direction as pixels there are in
an image frame. Initially all accumulators are empty; that is to say, their charge is
zero. When a jump between grey level bands occurs at a pixel, the charge unit (accu-
mulator) of the permanency memory at the pixel’s position is completely charged.
After the complete charge, each unit of the permanency memory goes decrementing
with time (in a frame-by-frame basis) down to reaching the minimum charge value,
while no motion is detected, or it is completely recharged, if motion is detected again.
Figure 1 shows all these issues. Figure 1a and 1b show two images of a monocular
sequence. The advance of a car may be appreciated, as well as a more slight move-
ment of a pedestrian. In figure 1c you may observe the effect of these moving objects
on the permanence memory.
Fig. 1. Permanency effect: (a) one image of a sequence, (b) same perspective after some sec-
onds, (c) motion trails as represented on the permanence memory.
The difference between a quick object as the car, which is leaving a very long mo-
tion trail (from dark grey to white), and a pedestrian whose velocity is clearly slower
and whose motion trail is nearly unappreciable with respect to the cars one, is pre-
sented. Thus, permanency memories enable representing the motion history of the
frames that form the image sequence, that is to say, there is segmentation from the
motion of the objects present in the scene.
Now, motion-based segmentation facilitates the correspondence analysis. Indeed,
motion trails obtained through the permanency memories charge units are used to
analyze the disparity between the objects in the stereo pair. The set of all disparities
between two images of a stereo pair is denominated the disparity map. The key idea is
that a moving object causes two identical trails to appear in epipolar lines of the per-
manency stereo memories. The only difference relies in their relative positions, af-
fected by the disparity of the object at each moment.
Looking at figure 2 it is possible to analyze the motion of each one of the three ob-
jects present in the permanency memories from their motion trails. This initial analy-
sis is independent of the epipolar constraint studied. You may observe that object “a”,
which has a long trail and has his maximum charge towards the left, is advancing to
the left at a high speed. Object “b”, with a shorter trail, is also advancing towards the
same direction but at a slower velocity. Finally, object “c”, whose trail is inverted in
horizontal, is moving to the right at a medium velocity, as shown by its trail.
Also from figure 2, but now comparing between the motion trails in both epipolar
lines, disparity is analyzed. Motion trail of object “b” presents a null disparity. There-
fore, we can conclude that this trail corresponds to an object that is far away from the
cameras. Remember that due to our parallel cameras configuration, pixels with a null
disparity are located in the infinite. Object “a” has a little greater disparity. Finally,
object “c“ offers the greatest disparity.
Left perma-
nency memory
Right perma-
nency memory
a b c
Fig. 2. Disparity of permanency memories.
The generalization to global analysis on complete stereo images consists in totally

superimposing the two permanency stereo memories under study, and not only their
epipolar lines. One of the memories will be displaced over the other looking for mo-
tion trails that coincide in both x and y directions. Once the displacement where the
coincidence of pixels of motion trails is maximum in size has been calculated, this
value is assigned the disparity value. By means of this functionality, for all possible
displacements of one permanency memory over the other, the correspondences be-
tween motion trails are checked and the disparities are assigned.
In order to test our algorithms, the scenario called “IndoorZoom” has been used.
Figure 3 shows the result some of the more representative results of applying our
algorithms to the “IndoorZoom” scenario. In column (a) some input images of the
right camera are shown, in column (b) the segmentation in grey level bands may be
appreciated, in column (c) motion information as represented in the right permanency
memory is offered, and in column (d) the final output, that is to say, the scene depth
as detected by the cameras, is presented.
Acknowledgements
This work is supported in part by the Spanish CICYT TIN2004-07661-C02-01 and

TIN2004-07661-C02-02 grants.
(a) (b) (c) (d)
Fig. 3. Results for “IndoorZoom” scenario
References
1. Brown, M. Z., Burschka, D. & Hager, G. D., “Advances in Computational Stereo”, IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 8, (2003).
2. Koenderink, J.A. & van Doorn, A.J., “Geometry of binocular vision and a model for stere-
opsis”, Biological Cibernetics, vol. 21, (1976): 29-35.
3. Wildes, R. P., "Direct recovery of three-dimensional scene geometry from binocular stereo
disparity", IEEE Transactions on Pattern Analisis and Machine Intelligence, vol. 13, no. 8,
(1991): 761-774.
4. Wilson, R. & Knutsson, H., “A multiresolution stereopsis algorithm based on Gabor repre-
sentation”, Proceedings of the IEE International Conference on Image Processing and Ap-
plications, (1989): 19-22.
5. Gutiérrez, S. & Marroquín, J.L., “Robust approach for disparity estimation in stereo vision”,
Image and Vision Computing, vol. 22, no. 3, (2004): 183-195.
6. Fernández-Caballero, A., Fernández, M.A., Mira, J., Delgado, A.E., “Spatio-temporal shape
building from image sequences using lateral interaction in accumulative computation”, Pat-
tern Recognition, vol. 36, no. 5, (2003): 1131-1142.
7. Fernández, M.A., Fernández-Caballero, A., López, M.T., Mira, J., “Length-speed ratio
(LSR) as a characteristic for moving elements real-time classification”, Real-Time Imaging,
vol. 9, (2003): 49-59.
8. Fernández-Caballero, A., Mira, J., Delgado, A.E., Fernández, M.A., “Lateral interaction in
accumulative computation: A model for motion detection”, Neurocomputing, vol. 50,
(2003): 341-364.
Pedestrian Detection for Intelligent Vehicles based on
Active Contour Models and Stereo Vision
C. Hilario, J. M. Collado, J. Mª Armingol, A. de la Escalera
Intelligent Systems Lab. Universidad Carlos III de Madrid. Leganes, Madrid, 28911. Spain
{chilario, jcollado, armingol, escalera}@ing.uc3m.es
Extended Abstract
In last years, the European Union, Member States and the automotive industry have concen-
trated their efforts on improving road safety [1][2]. Avoiding accidents due to human errors is
one of the main objectives of Advanced Driver Assistance Systems (ADAS). European traffic
accidents figures show a very large number of collisions between pedestrians and vehicles.
Every year, more than 200,000 pedestrians are injured, with over 6,000 killed. In spite of the
incredibly high number of victims, the protection of pedestrians has received little attention
by the research community [3]. The projects that have dealt with this problem are quite re-
cent, as has been pointed out at the Fifth Framework Programme [1].
The algorithm proposed in this paper, relies on the information provided by computer vi-
sion for the development of a pedestrian protection system [4]. The use of visual information
is a promising approach to cope with the different appearances of pedestrians and changes of
illumination in a cluttered environment. Although other sensors as lasers or radars can detect
pedestrians, vision is the unique that can comprehend their motion and predict their move-
ments. An active contour model is used to segment pedestrians from images taken by an on-
board camera. A stereo module is integrated to guide the location of active contours as they
are very sensitive to their initial position. Stereo vision is used to generate a disparity map of
the scene. In order to filter the disparity map, distance measures are employed and vertical
symmetries are looked for in the filtered area, avoiding a complete scan of the image. Those
regions with a high vertical symmetry are potential candidates for an active contour initializa-
tion.
Fig. 1. [Left]A detail of the Stereo Vision System used for the stereo module. [Right] On-
board PCs for the visual information processing.
Fig. 2. [Top-left] A right image taken by the Stereo System. [Top-right] Disparity map, [Bot-
tom-left] vertical symmetries and [Top-right] vertical edges of that image.
Active contours or “snakes” were proposed by Kass et al [5]. Its ability to extract contours
even in presence of gaps or occlusions, jointed to its dynamic behavior, makes this approach
adequate for non-rigid objects detection and tracking. The deformable model developed ex-
tends the greedy algorithm of Williams and Shah [6]. Their model maintains the points in the
snake more evenly spaced than Kass [5]. Moreover, the discretization of the model proposed
is controlled by inserting or removing points whenever the distance between adjacent points
became too large or too close. This way, the natural tendency to shrink of the snakes is com-
pensated. Several measures have been employed to avoid shrinkage, as controlling the
amount of stretching and bending of the contour locally or modifying the internal energy
formulation.
On the other hand, a new internal force is included to control the shape of the deformable
model. This regularizing force prevents the shrinking effect of the snake, as it is based on
high degrees of smoothness of the contour. It looks for convex and concave regions in the im-
age which are prone to correspond to the head and feet of a pedestrian. For the external
forces, a new potential field which smoothly extends over a long distance is defined. In order
to mitigate the influence of other contours with the same features as the ones rewarded by the
model, more specific knowledge about the target object is added. As pedestrians have a
strong vertical symmetry, the algorithm looks for points in the image that being symmetric,
are near enough from its vertical axis. Demanding furthermore that points being symmetric
and near a vertical axis must belong to a vertical edge, results are improved. Besides this po-
tential field, the image gradient and distances to vertical borders are also considered.
Once the snake is initialized on an object contour in the first frame, it will automatically
track the contour from frame to frame. This approach requires that the deformation and
movement of an object between frames is small. Otherwise, the snake could get stuck to other
contours or edges near the optimal. Some points in the snake are still prone to errors, like get-
ting trapped into the shadow of the pedestrian. This is solved applying restrictions to the re-
gion of search or assigning more stiffness to the model.
Initial Final
position snake
(a) (b) (c)
(d) (e) (f)
Fig. 3. [Top] Detail of a frame. In gray, the active contour from the previous frame is de-
picted. In white, the extracted contour. [Bottom] From (a) to (f) a complete sequence of
frames is shown.
This algorithm has been tested on images taken by a trinocular camera mounted on a
vehicle. A detail of the on-board vision system is shown in (Fig. 1). Some of the ob-
tained results are exposed in (Fig. 3).
Acknowledgements
This research is supported by the CICYT through the project TRA2004-07441-C03-01.
References
1 Information Society Technologies for Transport and Mobility. Achievements and Ongoing
Projects from the Fifth Framework Programme: Office for Official Publications of the
European Communities, 2003. <http://europa.eu.int>.
2 McDonald, J.; Markham, C.; McLoughlin, S. “Selected problems in automated vehicle

guidance”. Signals and Systems Group, 2001.
3 Gavrila D.M.; Kunert M.; Lages U. “A multi-sensor approach for the protection of vulner-
able traffic participants-the PROTECTOR project”. IEEE Instrumentation and Measure-
ment Technology Conference. 2001.
4 J. M. Collado, C. Hilario, a. de la Escalera, and J. M. Armingol, “Model based vehicle de-

tection for intelligent vehicles”, IEEE Intelligent Vehicles Symposium. pp. 572-577, 2004.
5 Kass, M.; Witkin, A.; Terzopoulos D. “Snakes: Active Contour Models”. Int. J. Comp.
Vision, 1(4), pp. 321-331, 1988.
6 Williams, D.J.; Shah, M. “A Fast Algorithm for Active Contours and Curvature Estima-
tion”. CVGIP: Image Understanding, 55(1), pp. 14-26, 1992.
Fast Road Sign Detection using Hough Transform for
Assisted Driving of Road Vehicles
Miguel Ángel García, Miguel Ángel Sotelo, Ernesto Martín Gorostiza
Department of Electronics, University of Alcalá, Alcalá de Henares, 28871

Email: garrido@depeca.uah.es, michael@depeca.uah.es, ernesto@depeca.uah.es
Abstract. A system for real-time traffic sign detection is described in this

paper. The system uses restricted Hough Transform to detect circular traffic
sings. Some results obtained on a set of real road images are presented in order
to illustrate the robustness of the detection system.
1 Introduction
Traffic sign detection and recognition has experimented increasing research interest in
the last times, this is due to the importance of improve safety in road vehicles. Drivers
sometimes miss signs because of distractions or lack of concentration. Our system
must alert the driver using voiced warning. There are four types of traffic signs in the
traffic code: prohibition, warning, obligation and informative. The most important
traffic signs are prohibition ones; therefore they have priority to be detected in this
work.
To detect a traffic sign in an image, the algorithm follows these steps:
1. Obtaining edge image using Canny method [1].
2. Looking for contours into the edge image.
3. Contours-filtering using aspect constraints.
4. Applying Hough transform [2] to filtered contours.
2 Contours Information
The method used for edge detection is Canny method; this method has been extremely
influential in many applications. Numerous implementations of edge-detectors based
on Canny’s idea have been developed. Canny described a method for generating an
edge-detector using an optimization approach and showed how to use the technique to
generate a robust-edges-detector. Canny’s method preserves contours, what is very
important for detecting traffic sign using shape information, because they are closed
contours. The contours are filtered using aspect constraints. In this sense, circular
contour has similar width and height, prohibition traffic signs are circular, and
therefore their contours will fulfill the aspect constraints. The Hough Transform is
applied to filtered contours, thus the computation-time is reduced. The search area of
traffic signs is focused in the right side of the image as it is showed in Fig. 1.
Fig. 1 Sequence of real road image, filtered edge image into the area of search and Canny
image.
3 Hough Transform
The classical Hough algorithm [3] can be easily extended to find any curves in an
image that can be expressed analytically in the form f(x, p) = 0 [4]. Here, x is a point
in the domain of the image and p is a parameter vector. We use the Hough algorithm
for circle location. A circle in the xy-plane with center (X, Y) and radius r can be
expressed as:
f(x, p) = (x - X)2 + (y - Y)2 - r2 = 0 (1)
Where the parameter space, p = (X, Y, r), must be quantized. The accumulator matrix
a is the representation of the quantized parameter space. For circle detection the
accumulator a will be a three-dimensional matrix with all entries initially set to 0. The
entry a(Xr, Ys, rt) is incremented by 1 for every feature point (xi, yi ), in the image-
domain, contained in the circumference with centre (Xr, Ys) and radius rt, with a
precision margin by ε, used to compensate quantization error when digitize the image
[5]. This condition is expresed:
|(Xr – xi)2 + (Ys – yi)2 – rt2| < ε (2)
In order to reduce the computation time we use contour information, so, every feature
point (xi, yi ) is searched inside a restricted area in every filtered closed-contour.
3 Results and Conclusion
Fig. 2 Sequence of real road images under adverse weather conditions where speed limit is
detected.
An algorithm for real-time detection of traffic signs is carried out based on Hough
Transform. The algorithm has been empirically tested under adverse weather
conditions. We detect the road sign when its closed contour increases the accumulator
matrix a over one threshold, this condition is easy to fulfill for circular shapes. If
several circular shapes are detected it will not be a problem since a neural network
will make a fine classification and validation of the detected road signs. As a future
work this method for circular sings detection can be easly extended to searching
triangular traffic sings using the Hough Transform for lines expressed as:
x cos(β) + y sin(β) = r (3)
A triangular traffic sing is delimitated by three straight lines. A straight line is
expressed analytically in the form f(x, p) = 0, where p(r, β) is the parameter vector.
The accumulator matrix a will be a two-dimensional matrix. The rest of this method is
very similar to circular traffic-signs detection. Our system works between 5 and 50
frames/s, depending of the amount of shapes detected.
References
1. J. Canny. “A Computional approach to Edge-Detection,” IEEE Transactions on pattern

Analysis and Machine Intelligence, vol 8, pp. 679-700, 1986.
2. R. Duda and P. Hart, “Use of the Hough transform to detect lines and curves in pictures,”
Communications of the ACM, vol. 15, no. 1, pp. 11-15, 1972.
3. P. Hough, “Method and means for recognizing complex patterns,” Dec. 18 1962. U.S. Patent
3,069,654.
4. D. Ballard, “Generalizing the Hough transform to detect arbitrary shapes,” Pattern
Recognition, vol. 13, no. 2, pp. 111-122, 1981.
5. S. D. Shapiro, “Properties of transforms for the detection of curves in noisy image,”
Computer Graphics and Image Processing, vol. 8, pp. 219-236, 1978.
Advances in Robotics
P. Kopacek
Intelligent Handling and Robotics - IHRT

Vienna University of Technology, Austria
kopacek@ihrt.tuwien.ac.at
Extended Abstract
The field of robotics is one of the most innovative in the last decade. Conventional
industrial robots from the late 70´s are now only a tool on the production level. One of
the oldest dreams of the robotic community – intelligent, mobile, cooperative as well
as humanoid robots – starts to become reality not only because of the rapid
development of “external” sensors and micro- and nanotechnology.
External sensors (e.g. visual, auditive, force-torque…) combined with micro
drives, embedded systems,… offer intelligent robots the possibility to see, hear,
speak, feel, smell like humans. Compared with conventional, unintelligent, industrial
robots, intelligent robots fulfill new, innovative tasks in new application areas.
There are three “starting” points for the development of intelligent robots:
Conventional, stationary industrial robots; mobile, unintelligent platforms (robots)
and walking machines.
Stationary industrial robots equipped with external sensors are used today for
assembly and disassembly operations, fluelling cars, cleaning of buildings and
airplanes, humanitarian demining ... and have been the first “intelligent” robots.
Mobile platforms with external sensors are available since some years and cover a
broad application field. The core of each robot is an intelligent mobile platform with
an on-board PC. On this platform, various devices, like arms, grippers, transportation
equipment, etc., can be attached. Communication between the „onboard PC“ and the
„supervisory PC“ is carried out by radio-based networks - communication with the
environment can be accomplished by voice, beep or bell.
The application of “agents” in production automation is closely related to
“holons”. Following the results of TC5 of the worldwide ims initiative a holon is an
identifiable part of a system that has a unique identity, yet is made of subordinate
parts and in turn is part of a larger whole. This similar to the usual „fuzzy“ definition
of an agent. The features of a „technical“ agent are:
• Ability to optimize one or more processes simultaneously
• Autonomy – an agent can made own decisions to stabilize to current state and to
improve this state to reach to optimum of the whole system.
• Communication and interaction – for reaching this optimum each agent must
interact with the other agents in the MAS .
The ultimate target to be reached would be a robot that possesses capability
approaching that of human beings - autonomous robot agents. Leaving such an ideal
robot as a goal for the future, intermediate robots that only satisfy a limited selection
of the most requisite functions should still find good use in human society. Among the
faculties cited above, mobility is the most indispensable feature for a service robot.
Walking machines or mechanisms are well known since some decades. Usually
they have 4 to 6 legs (multiped) and only in some cases 2 legs (biped) – walking on
two legs is from the view of control engineering a very complex (nonlinear) stability
problem. Biped walking machines equipped with external sensors are the basis for
“humanoid” robots.
It was an old dream to have a personal robot looking like a human. Main features
of a humanoid robot are
• biped walking
• voice communication – speech recognition
• facial communication
The main feature of a real human is the two legged movement and the two legged
way of walking. Much research has been conducted on biped walking robots because
of their greater potential mobility. On the other side they are relatively unstable and
difficult to control in terms of posture and motion.
Currently there are worldwide two categories of two legged humanoid robots
available:
“Professional” humanoid robots developed by large companies with a huge
amount of research capacities. Examples are: the Honda robots ( P1, P2, P3, ASIMO)
– with the idea to assist humans in everyday working, the SONY robots ( SDRX –
3,4,5) – with the background to serve mostly for entertainment, leisure and hobby or
in the future as personal robots.
“Research” humanoid robots: There a lot of such robots currently available or in
the development stage e.g. approximately worldwide more than 500 University
institutes and research centres are active in this field. The robots of this category a
usually prototypes developed by computer scientists to implement methods of AI,
image processing, theoretical scientists from mechanics implementing and testing
walking mechanisms, control scientists to implement new control strategies, social
scientists to implement human machine interfaces (HMI) for an efficient
communication between humans and humanoid robots.
In addition these intelligent robots – especially mobile platforms and humanoid
robots - are able to work together on a common task in a cooperative way . The goal
are so called “Multi Agent Systems – MAS”. MAS consist of a distinct number of
robots (agents), equipped with different arms, lifts, tools, gripping devices, ... and a
host computer. A MAS has to carry out a whole task e.g. assemble a car. The host
computer divides the whole task in a number of different subtasks (e.g. assembling of
wheels, windows, brakes,...) as long as all this subtasks can be carried out by at least
one agent. The agents will fulfill their subtasks in a cooperative way until the whole
task is solved.
In industry intelligent robots will work together with humans in a cooperative way
on a common working place.
One of the newest application areas of service robots is the field of entertainment,
leisure and hobby because people have more and more free time. In addition modern
information technologies lead to loneliness of the humans (teleworking, telebanking,
teleshopping,....). Therefore service robots will become a real “partner” of humans in
the nearest future. One dream of the scientists is the “personal” robot. In 5, 10 or 15
years everybody should have at least one of such a robot. Because the term personal
robot is derived from personal computer the prices should be equal.
In the paper the present state will be discussed, selected application described and
an outlook on future developments will be given.
Current and Future Trends and Challenges in Robot
Soccer
Norman Weiss, Bernd Reusch, Lars Hildebrand
University of Dortmund, Chair Computer Science I, Otto-Hahn-Str. 16,

44227 Dortmund, Germany
{Bernd.Reusch, Lars.Hildebrand, Norman Weiss}@uni-dortmund.de
Extended Abstract
Robot soccer has evolved into a very dynamic and competitive field within the last
few years. It provides excellent research and benchmarking opportunities in a diverse
variety of fields, e.g. multi-agent systems, robot control, sensor fusion, intelligent
control, communication, image processing, mechatronics and artificial intelligence.
Many different robot soccer leagues now exist. The scope ranges from truly tiny,
centrally controlled robots in FIRA NaroSot (4.5 x 4.5 x 5 cm) to humanoid, fully
autonomous robots in humanoid leagues (up to 180 cm in height). At the same time,
the game play of the different leagues can be everywhere between highly autonomous
and research-oriented to high-speed and entertainment-oriented.
The robot soccer league most strongly dedicated to entertainment and edutainment
is currently FIRA MiroSot. This paper will present the current state of robot soccer
within FIRA MiroSot and upcoming trends and challenges.
In the FIRA MiroSot league, the size of a robot is limited to 7.5 x 7.5 x 7.5 cm, which
is very small considering the tasks it has to fulfill. The league started in 1996 with 3
vs 3 robots playing on a field sized 150 x 130 cm. To get closer to “real” soccer, the
number of players on the field slowly grew (5 vs 5, 7 vs 7) along with the field. It has
now reached 11 vs 11 robots on a 440 x 220 cm field.
During the early days of the league, the robots’ mechanical designs, their control
and the complex demand of the strategy were limiting factors. Since then, it has
evolved into a very dynamic, high-speed game with robots reaching speeds of up to
3.5 m/s during game play. Alongside with now refined strategies, this provides a very
entertaining experience to spectators.
Due to the very small size of the robots, they must be supported by a host computer
(usually an off-the-shelf PC) which receives a picture of the field from a camera
mounted about 2.5 m above the field. The host is responsible for image processing
and strategic decisions. It transmits – via a radio link– movement information to the
robots on the field, which they execute, thereby closing the control cycle.
While some problems robot soccer posed have been solved in the last few years
[1], many still exist. At the same time, constant change in the league adds new prob-
lems – especially the change to 11 vs 11 robots has provided many new challenges to
the teams, which will be presented in the following.
The 11-vs-11-field is sized 440 x 220 cm. The cameras are currently mounted at 2.5
m above the field, which means that strong fish eye lenses are necessary to get an im-
age of the entire field. That strong fish eye effect is very undesirable, since it impedes
exact position calculations of the robots. Raising the camera to a higher level is not
possible, since this would rule out too many locations for playing. That means that
most teams playing the 11-vs-11-league must revert to dual camera approaches,
where every camera records one half of the field [2]. Insights into the challenges and
solutions to this problem will be presented.
All FIRA MiroSot robots are controlled by the host computer via a radio link. Cur-
rently, most teams use BiM series radio modules manufactured by the English com-
pany Radiometrix [3]. These transmit in the 433 or 869 MHz ISM bands with speci-
fied gross data rates of up to 160 kbps. Unfortunately, due to various reasons, the net
data rate will be around 35 kbps in this application, with the modules being halfdu-
plex only.
With 22 robots on the field playing robot soccer, the limits of the Radiometrix
modules are now reached. New solutions like differently designed radio hardware
transmitting in different frequency bands as well as updated and more stable protocols
are needed. Selected alternatives to the current communication links will be pre-
sented.
Originally, only 6 robots (3 vs 3) were playing during a FIRA MiroSot game. This de-
sign put emphasis on control and vision problems, because the matter of strategy was
deliberately kept simple. One goalkeeper and two robots actually playing the game
did not allow for any refined strategic moves. This has strongly changed during the
last few years. Strategy design will become much more difficult within FIRA MiroSot
because of the recent extension to 22 (11 vs 11) robots on the field [4]. Simple ap-
proaches as outlined above will not suffice and much more complex strategies will be
needed that get much closer to “human” soccer. Complex team based behaviours like
passes that require a highly sophisticated strategy and coordination will be of advan-
tage to the team mastering them.
Some insight into formerly and currently possible strategy design(s) will be pre-
sented.
References
1. Kim, J.-H., Kim, D.-H., Kim, Y.-J., Seow, K.-T.: Springer Tracts in Advanced Robotics,
Vol. 11: Soccer Robotics. Springer Verlag, Berlin Heidelberg New York (2004)
2. Weiss, N., Jesse, N.: Towards Local Vision in Centralized Robot Soccer Leagues: A Robust
And Flexible Vision System Also Allowing Varying Degrees Of Robot Autonomy. Proc.
FIRA World Congress 2004, Busan, Korea (2004)
3. 869/914MHz High Speed FM Radio Transceiver Module BiM3 Data Sheet. Radiometrix
Ltd, Watford, UK (2004)
4. Hildebrand, L., Michalski, C., Valentin, H., Wickrath, M.: Strategy Implementation For
Mobile Robots Using The Pipes & Filters Architecture. Proc. FIRA World Congress 2003,
Vienna, Austria. (2003)
Strategy and communication in robotic soccer game
Bohumil Horak1, Marek Obitko2, Jan Smid3, and Vaclav Snasel1

1
Department of Computer Science, FEI, VSB - Technical Univerzity Ostrava,
17. listopadu 15, 708 33, Ostrava-Poruba, Czech Republic
{Bohumil.Horak, Vaclav.Snasel}@vsb.cz
2
Gerstner Laboratory, Department of Cybernetics,
Faculty of Electrical Engineering, Czech Technical University
Technicka 2, 166 27 Prague, Czech Republic
obitko@labe.felk.cvut.cz
3
Computer Science Department, Morgan State University,
Baltimore MD, USA
jsmid@jewel.morgan.edu
Abstract. We describe the key components of our game strategy for robot
soccer [3] in detail. The game can be represented as a trajectory in so-called the
virtual grid. The virtual grid generally allows us to reduce data volume for easy
description of player motion and subsequently for controlling the game or for
learning game strategies. The natural coordinate system is provided by accurate
optical sensing of the subject position using lens for optical transformations and
the CCD camera. This natural coordinate system can be easily mapped to a
virtual grid.
Using the virtual grid, it is possible to describe the position and direction of the robot
using an alphanumeric description see [1,2]. The primary virtual grid can be divided
to (2, 4, 8, . . .) parts, which creates secondary virtual grid.
The tercial grid — strategy grid — depends on the partition of the game field (the
left-right wing, the central field, and transversely the attack-defense field and the
central field). In the discrete frame samples it is possible to study movements and
movement strategies of the robots.
Fig. 1. Alphanumeric representation of robot position and movement using virtual grid.
The game strategy can be dynamically changed based on the game progress (i.e. the
history and the current position of the players and the ball). The game progress can
be divided in time into the following three ground playing classes (GPC):
• GPC of game opening (GPCO)
• GPC of movements in game site (GPCS)
• GPC of game end (GPCE).
The game progress, especially in the GPCS class, can be also divided into the
following two game playing situations (GPS):
• GPS of attack (GPSA). The interactions of simple behaviors cause the
robots to fall into a V-formation where the ball is in motion roughly towards
the opponent’s goal.
• GPS of defense (GPSD). When the ball is not moving roughly towards the
opponent’s goal, the robots move around it to form an effective barrier and
to be in a good position for recovery.
Each GPC has its own movement rules. The classes GPCO and GPCE consist of
finite number of possible movements that are determined by initial positions of
players and the ball. The class GPCS has virtually unlimited number of possible
movements. The movements are determined by the current game situation (GPS) and
by the appropriate global game strategy (in next GGS). The movement of the
particular robot is determined by the current game class and situation, and also by the
robot role. For example, the goalkeeper’s task is to prevent the opponent to score a
goal. His movements are in most cases limited along the goalmouth near of goal line.
The preferred movements are in goal line direction. The preference of these
movements comes from the particular GGS, where the goalkeeper prevents to score a
goal in the way of moving in the position between the central goal point and the ball
(or the expected ball position). The preference of other movement directions is
created using GPSA, where the movements of goalkeeper secure kicking the ball
from the defense zone.
Each strategy is stored in one file and currently consists of about 50 basic rules.
Furthermore the file contains following metadata:
• Information about the name of strategy
• the algorithm to strategy choosing
• the author responsible for current strategy
• the date of last modification
• the size of strategic grid
• strategic rules
Each rule consists of five records:
• The rule ID and description (e.g. Rule 1 "desc"),
• the coordinates of our players in strategic grid (e.g. .Mine a5 a5 a5 a5 a5),
• the coordinates of opponent's players in strategic or virtual grid
(e.g. Opponent k5 j2 i5 h7 i8),
• the ball coordinates in virtual or strategic grid (e.g. .Ball j8)
• strategic or virtual grid positions of the move (.Move a5 i4 g5 i6 j8).
In current algorithm, the .Mine coordinates are not important for movement rule
selection, but they can be used in the future. We believe that the file system for
strategies is an advantage. From observation of opponent's strategy a new set of rules
can be written, without necessity of program code modification. Furthermore, there
is a possibility of automatic strategy (movement) extraction from running game.
.Strategy test 1
.Author "Vaclav Snasel"
.Date "1.5.2004"
.Size 10 8
.Rule 1 "Attack1"
.Mine a4 b3 c1 d1 d4
.opponent a2 b3 c2 d3 d4
.Ball a1
.Move a4 b3 c1 d1 d4
.Rule 2 "Attack2"
.Mine a4 b3 c11 d1 c10
.Opponent a2 b3 c2 d3 a1
.Ball a1
.Move a4 b3 c1 d1 a11
There exist two main criteria in the Strategy selection process. The selection depends
on opponent’s coordinates and ball position. The strategy file contains rules,
describing three possible formations suggesting danger of current game situation.
The opponent’s team could be in offensive, neutral or defensive formations.
Furthermore, we need to weigh up the ball position risk. Generally, opponent is not
dangerous if the ball is near his goal. The chosen rule has minimal strategic grid
distance from current configuration of players and ball.
Optimal movements of our robots are calculated by applying minimal distance from
strategic grid position. The goalkeeper and attacking player, whose distance is
closest to the ball are excluded from strategic movement and their new position is
calculated in exact coordinates
Acknowledgements
The Czech Academy of Sciences partially supported this project within the grant
“Strategic control of the systems with multiagents“.
References
1. Bernatik, R., Horak, B., Kovar, P.: Quick image recognize algorithms. In: Proceeding
International workshop Robot-Multi-Agent-Systems R-MAS 2001. VSB-TU Ostrava
2001, Czech Republic 2001 p.53-58.
2. Horak B., Obitko M., Smid J., Snasel V.: Communication in Robotic Soccer Game.
Communications in Computing 2004: Las Vegas, p. 295-301.
3. Kim, J.; Kim, D.; Kim, Y., Seow, K.: Soccer Robotics (Springer Tracts in Advanced
Robotics), Springer-Verlag, 2004.
Rete algorithm applied to robotic soccer
M. Palomo, F.J. Martı́n–Mateos & J.A. Alonso
Computational Logic Group

Dept. of Computer Science and Artificial Intelligence, University of Seville
E.T.S.I. Informática, Avda. Reina Mercedes, s/n. 41012 Sevilla, Spain
E-mails: manuel.palomo@uca.es, {fmartin,jalonso}@cs.us.es
1 Introduction
Robotic soccer is one of the most interesting examples of multi-agent system.

In this environment, agents must be able to perform as a team to get a com-
mon long-term goal. They have to manage themselves in a real-time, non-
deterministic, partially known world. All this facing a team whose goal is the
opposite (both teams can’t fulfil their goals at the same time).
This paper is focused on simulated robotic soccer. The simulation system
allows to deal with high level problems instead of wasting time with low level
details. Scientists can apply automated reasoning techniques (off-line and on-
line). It also offers the advantages of software utilities: recording matches (and
replaying them later), playing some teams versus other teams (if their imple-
mentations are available), building teams with exactly the same resources, etc.
Anyway, under some constraints, direct implementation of simulated soccer al-
gorithms in real robots is possible.
2 Foundations
Our proposal is based on the use of Rete algorithm [1] to design a rule-based
expert system. Expert systems design is a branch of artificial intelligence spe-
cialized in development of systems that simulates expert human decision skills.
In particular, rule-based expert systems are composed of a set of facts (knowl-
edge base) and a set of if-then statements (rules). Facts represent information
known (or believed) about the world. And rules describe how that information
could change. The set of rules are fixed during all the execution of an expert
system, but facts can change. Every rule has two parts: the antecedent (a set
of conditions about facts) and the consequent (a set of actions that modify the
knowledge base adding or deleting facts). One of the key features of rule-based
expert systems is the possibility of increasing the knowledge (adding new rules)
without losing the information in the knowledge base.
The expert systems work in an infinite loop. In each cycle they check the
antecedent of every rule. If all the conditions of one of them are true the rule is
activated (and placed on the agenda). The agenda is the list of all rules which
have their conditions satisfied and have not yet been executed. After checking
all the rules, one of the activated ones (that are on the agenda) is selected
to be triggered. The selection method is called conflict resolution strategy, and
depends on the implementation of the system. There are several of them, such as
assigning a priority to every rule, choosing the most recent activated, the most
times triggered, etc.
In every cycle of simulation the system has to check all the new set of facts
and rules that are activated. A direct implementation of an algorithm checking
all the conditions of every rules would be too inefficient.
Rete algorithm takes advantage of two facts [4]. First, most of the facts of
an expert system don’t change from one cycle to another. Thus, it doesn’t check
all the rules every cycle, but it remembers past facts test results, so only new
facts are tested. And second, as several rules can share part of their antecedents
a network is created to minimize the number of tests to be made in each cycle.
It all results in a more efficient algorithm (on the average). The figure 1 shows
an example of optimization of two rules.
?X ?Y ?Z ?X ?Y ?X ?Y ?Z
+ + +
+ +
Trigger rule 1 Trigger rule 2 Trigger rule 1 Trigger rule 2
Fig. 1. Example of two rules optimized creating a network sharing conditions.
This algorithm could be used to implement a robotic soccer team composed of

robots programmed as expert systems. In particular, the architecture proposed is
for a team for Robocup Soccer Server [3]. Robocup Soccer Server is a system that
simulates robotic football matches in a 2D field. It’s supported by the Robocup
Federation and it’s widely used. It works with a C/S architecture. In each match
there is one server simulating everything concerning the match, and 22 clients
(11 for each team), each one controlling one player. Clients are autonomous, and
they can only communicate through the server. It implements a low bandwidth,
unreliable communication channel.
The server simulates the world in steps of 100 ms. Every step it accepts one
action (like kick or turn) from each client, and simulates them all (applying real
physics and football rules like decay, off-side rule, etc). Part of the result of this
simulation is sent to the clients every cycle. That information is of three kinds:
aural (what a player hears), visual (what a player sees) and physical (what
a player knows about himself, like stamina, speed and so). Every player only
receives the information it senses: messages heard, objects seen and physical
information.
3 Proposed architecture
The architecture modifies the proposal of Peter Stone in [2]. It, basically, uses
some information common to all the players in a team (called locker-room agree-
ments) and the information received from the world (partial information) to
update the internal state of the agent (Figure 2). With those internal beliefs
a directed acyclic graph (known as external behaviors) selects the action to be
made (an example is shown in figure 3).
TEAM MEMBER AGENT ARCHITECTURE
Agreements Internal
state
World state External

behaviors
Information Actions
Real world
Fig. 2. Simplified internal architecture of an agent.
We propose two modifications of this model. First, all the information re-
ceived from the world is stored as low-level facts. But it’s also generalized as
high-level information relative to objects in the field, in such a way that small
changes in the world won’t lead to changes in the high-level perception of it. Of
course, this generalization must be according to soccer concepts: for example, a
set of positions of players can be generalized in a “offside” fact.
And the second is the use of Rete algorithm to choose the action to be per-
formed. In this case the antecedents of the rules must check high-level properties
(so the algorithm won’t have to be re-testing a lot of rules every cycle), but the
consequent (that chooses the action to be made) can use low-level facts in order
Behavior Face Ball()
...
Behavior Play Soccer() Behavior Handle Ball(args)

...
If (Ball Lost) Face Ball()
If (Ball known AND Chasing) Handle Ball(args1)
If (Ball known AND Not Chasing) Passive Offensive(args2) Behavior Passive Offensive(args)
If (Communicate Flag Set) Communicate() ...
Behavior Communicate()
...
Fig. 3. Example of external behavior as implemented in [2]. It’s replaced by a set of

rules optimized in a network (like in figure 1).
to calculate the best action according to accurate information. So the directed

acyclic graph is substituted by a set of rules managed with Rete algorithm. Ac-
cording to offside example, if “offside” fact exists in the knowledge base, rules
as “pass forward” can’t be activated. And a light movement of the player to
another offside position won’t lead to the activation of any of those rules. But if
the movement avoids offside, the rules could be activated.
References
1. Forgy, C. Rete: A Fast Algorithm for the Many Pattern/Many Object Pattern
Match Problem. The Journal of Artificial Intelligence, Vol. 19, 1982, pp. 17-37.
2. Stone, P. Layered Learning in Multi-agent Systems. PhD thesis, School of Computer
Science, Carnegie Mellon University, 1998.
3. Robocup team. Soccer Server System. http://sserver.sourceforge.net.
4. Giarratano & Riley. Expert systems: principles and programming, third edition.
PWS Publishing Company, 1998.
Vision system design and accompanying problems in
robot soccer
Gregor Klančar, Drago Matko, Rihard Karba
University of Ljubljana, Faculty of Electrical Engineering,

Tržaška 25, SI-1000 Ljubljana, Slovenia
gregor.klancar@fe.uni-lj.si
Extended Abstract
There are many ways of detecting moving objects using color cameras. However, the
vision systems based on color information proved to be more simple, robust and faster
than most of other recognition methods as stated in [2], [8] and [11]. Sargent et al. [8]
developed a fast real-time vision system with the aid of a special hardware accelerated
system, which only makes sense when software optimizations or accelerations are not
possible. A more reliable vision tracking of moving objects can be achieved by using
robust statistics and probability distributions. A good example of the latter is given in
the color-based face tracking implemented by Bradski [1]. Bruce et al. [2] suggested a
fast vision system for mobile robots by means of efficient color segmentation and a
two-pass connected region determination algorithm. Another important contribution
to the robot soccer vision design was introduced by Wyeth et al. [11], with special
consideration given to the robustness of varying playground illumination conditions.
This paper presents a vision system for mobile robots tracking with the stress given
to the accompanying problems in vision system design such as noise filtering, camera
lens distortion and non-uniform illumination. The basic operation of a global vision
system based on color information is introduced. It is used for mobile robots control
in well-defined and fast dynamic robot soccer game. Therefore simple and efficient
algorithms should be used to optimize the processing time needed for robots
estimation.
Further on an approach to improve an already existing vision system performance
under bad operating conditions is presented. Some fundamentals and solutions to
accompanying problems in vision system design for mobile robot tracking are
presented. Besides methods for filtering and improvement of identified noisy data the
two main factors which deteriorate the performance are dealt with, namely, non-
uniform illumination and camera lens distortion. For the former the problem area and
its origins are focused on and a solution for its compensation by applying
multiplicative component defined by illumination plain is given. The latter consists of
two steps. In the first, radial lens distortion fundamentals are discussed. The suggested
solution for its verification is realized by a geometry model of lens projection. The
second step covers the perspective distortion originating from the tilt of the camera.
For its correction an efficient and robust method of vanishing point detection is
applied. Both correction methods contribute to a vision system performance if
implemented in the appropriate manner.
Applicability of the presented approaches is confirmed on a robot soccer test bed.
Robot soccer is a fast dynamic game and therefore needs an efficient and robust
vision system. To improve the results of the robot soccer vision system the proposed
camera calibration and non-uniform illumination correction algorithm are
implemented. The lens correction method successfully corrects the distortion caused
by the camera lens, thus achieving a more accurate and precise estimation of the
object position. The illumination compensation improves robustness to irregular and
non-uniform illumination which is usually present in real conditions.
References
1. Bradski, G.R.: Computer Vision Face Tracking For Use in a Perceptual User Interface.
Intel Technology Journal. 2nd quarter (1998) 1-15.
2. Bruce, J., Balch, T., Veloso, M.: Fast and inexpensive color image segmentation for
interactive robots. In: Proceedings of IROS-2000. Takamatsu Japan (2000) 2061-2066.
3. Fangi, G., Gagliardini, G., Malinverni, E.S.: Photointerpretation and small scale
stereoplotting with digitally rectified photographs with geometrical constraints. In: CIPA
congres 2001. (2001) http://cipa.icomos.org/papers/2001-.htm.
4. Likar, B., Maintz, B.A., Viergever, M.A., Pernuš, F.: Retrospective shading correction
based on entropy minimization. Journal of Microscopy. 197 (2000) 285-295.
5. Pajdla, T., Werner, T., Hlavač, V.: Correcting Radial Lens Distortion without Knowleadge
of 3-D Structure. Technical report TR97-138. Faculty of Electrical Engineering, Czech
Technical University, Praha, Czech Republic (1997).
6. Perš, J.: Kovačič, S. Nonparametric, Model-Based Radial Lens Distortion Correction Using
Titled Camera Assumption. In: Wildenauer, H., Kropatsch, W. (Eds.): Proceedings of the
Computer Vision Winter Workshop 2002. Bad Aussee Austria (2002) 286-295.
7. Rother, C.: A new Approach for Vanishing Point Detection in Architectural Environments.
In: Proceedings of the 11th British Machine Vision Conference (BMVC'00). Bristol UK
(2000) 382-391.
8. Sargent, R., Bailey, B., Witty, C., Wright, A.: The importance of fast vision in winning the
First Micro-Robot World Cup Soccer Tournament. Robotics and Autonomous Systems. 21
(1997) 139-147.
9. Tsai, R.Y.: A versatile camera calibration technique for high-accuracy 3D machine vision
metrology using off-the-shelf TV cameras and lenses. IEEE Journal of Robotics and
Automation. 3(4) (1987) 323-344.
10. Tuytelaars, T., Proesmans, M., Gool, L.V.: The cascaded Hough transform. In: Proceedings
of International Conference on Image Processing (ICIP ‘97), Vol. 2. Washington (1997)
736-739.
11. Wyeth, G.F., Brown, B.: Robust Adaptive Vision for Robot Soccer. In: Billingsley, J. (ed.):
Mechatronics and Machine Vision in Practice. Research Studies Press UK (2000) 41-48.
12. Jetto, L., Longhi, S., Venturini, G.: Development and Experimental Validation of an
Adaptive Extended Kalman Filter for the Localization of Mobile Robots. IEEE
Transactions on Robotics and Automation. 15(2) (1999) 219-229.
Time-optimal collision avoidance of
automatically guided vehicles
Albert L. Schoute
University of Twente, Department of Computer Science

Postbox 217, 7500AE Enschede, Netherlands
a.l.schoute@utwente.nl
Extended abstract
Collision avoidance is a main issue in applications employing free-ranging, com-

puter-controlled vehicles. In this paper we look at the regulation of locally crossing
traffic. Collisions are typically avoided by changing speed and/or changing route,
but– to keep efficiency high– as minimal as possible.
If we look at humans going by foot in crowded areas or driving by car in urban
traffic, we observe a high degree of flexibility. The challenge is to reach the same
flexibility with computer controlled driver systems. Preliminary research on dynamic
ways of traffic control [1] has indicated that considerable improvements can be
reached in comparison to the more static approach of traditional zone claiming
methods, in particular in case of busy traffic.
When searching for the best solutions there are many criteria to consider: time de-
lay, throughput, total travel time, planned arrival time, energy consumption, com-
fort, et cetera. We will limit our selves to a single optimality criterion: minimal time
delay. Even then we are faced with a huge space-time resource allocation problem
[2].
In the paper, first we analyze the passage of two vehicles that drive on a conflict-
ing course as depicted in the left part of Fig. 1.
v1 = v2
progress of time
d2
trailing delay
v1 vehicle
Collision
states bumper
α
sticking
v2 blocking
d1 ∆
progress of leading vehicle

Fig. 1. Joint motion state diagram (shown on the right) of two vehicles crossing with angle α
(as shown on the left) with the leading vehicle having advantage ∆ = d2-d1 and the trailing
vehicle adjusting its speed (v2 ≤. v1).
Basically, the following options for collision avoidance exist: (1) speed adaptation,
(2) route deviation by one vehicle only, (3) route deviation by both vehicles and (4)
combined speed and route adjustment. In the case of speed adjustment generally only
the vehicle driving behind (i.e. the “trailing” vehicle) will slow down (or make a
temporary stop). The resulting delay depends on the passing strategy: a minimal
delay is obtained if the trailing vehicle immediately advances after the leading vehi-
cle (“bumper sticking” mode). For safety reasons one may require that the trailing
vehicle only proceed if its pathway is completely cleared (“blocking” mode). The
observed time delay can be visualized in a state diagram representing the joint mo-
tion (right part of Fig. 1). The indicated area of collision states depends on the vehi-
cle sizes and the crossing angle [5].
In Fig. 2 we show two ways of vehicle passing by route deviation. In the left draw-
ing only one vehicle makes a detour by changing its direction three times. In the
right drawing both vehicles make a detour along a (virtual) roundabout.
v2
v2
v1
v1
Fig. 2. Examples of unilateral and bilateral collision avoidance detours.

Under certain symmetrical assumptions the optimal avoidance maneuver of two
vehicles giving minimal mean delay has been determined. If conditions become
asymmetric due to unequal priorities or different destination distances, it becomes
more difficult to pinpoint the optimal solution.
An important aspect of the result obtained for two passing vehicles is that it indi-
cates an approach to guide multiple conflicting vehicles in passing each other effi-
ciently. By means of simulation experiments it has been shown in [2] that this ap-
proach leads to quite natural avoidance detours for all kinds of multiple robot cross-
ing patterns. The solution method very much contrasts to the usual approach of plan-
ning motion paths sequentially in some optimal priority order [4].
The shape of the optimal avoidance maneuver of two vehicles can be found by
global arguments. During an optimal avoidance detour the vehicles will come “in
close contact” somewhere (without colliding), otherwise a shorter detour is possible.
The fastest way to reach the first close contact configuration (the “meeting” state)
will be along a straight line from the starting points. The same holds for the fastest
way to leave the last close contact configuration (the “greeting” state) in order to
approach the destination. In between “meeting” and “greeting” the vehicles remain
in close contact. Otherwise they would make a relative movement that – given some
maximal speed - slows down reaching the greeting state.
As long as vehicles move along straight lines, their orientation will remain in-
variant. However, the detour implies orientation changes that in the end have to sum
to zero. In the case of equal distances from start to destination the vehicles will make
a symmetric and synchronized detour. To compensate for the orientation deviations
at the starting and destination points (with respect to the direct route) the vehicles –
being in close contact – will rotate simultaneously, i.e. move as a coupled pair along
a virtual roundabout as shown in Fig. 3.
α
β
“meeting state” “greeting state”
Fig. 3. Time optimal avoidance detour over equal distances.

The shape of the vehicles determines the size of the roundabout, as they must fit to
the roundabout. The placement of the roundabout is determined uniquely as the ve-
hicles must meet on the roundabout at the same time. As a consequence, the optimal
deviation angle β and the length of the detour route can be calculated. The delay is
found by comparing the detour route with the direct route.
If the travel distances to the destinations differ, deviation angles will also differ.
Hence, the turning to compensate for the orientation changes will not be equal. Mak-
ing different turns along a common roundabout is certainly a near-optimal solution.
However, it is not yet clear if this maneuver is also optimal.
An obvious property of the optimal avoidance detour is that it remains optimal at
any intermediate state as long as the destinations do not alter. Therefore, steering the
vehicles at any instant along the optimal deviation route will lead automatically to
the optimal passing maneuver. Consequently, reactive control can be applied to
reach optimal avoidance. Reactive control has the additional advantage that it is
stable with respect to errors. Moreover, it will adapt to changes of the destinations.
Hence, it can be applied fruitfully to vehicle control systems with dynamically chang-
ing goals like for instance robot soccer.
The optimal passing maneuver of two vehicles provides a strong “heuristic” to at-
tack the case of multiple controllable robot vehicles. For each pair of robots the op-
timal deviations necessary to avoid a collision are known. Of course, if more robots
are involved, different and sometimes incompatible avoidance courses have to be
satisfied. The best strategy will be to find some acceptable compromise, for example
taking the largest deviation needed. Clusters of colliding robots have to be identified
of which the deviations must be coordinated. The strategy comes down to finding the
best “virtual roundabout” that solves the mutual conflicts within the cluster. It has
been shown that applying the deviation control in an iterative, reactive manner does
work effectively in most cases [3].
Pure route deviation without speed adaptation could in case of multiple conflicts
lead to large, undesirable detours. Also avoiding a non-cooperative robot by unilat-
eral route deviation may be less favorable than lowering speed. Good heuristics for
the optimal mixture of route deviation and speed adaptation still have to be discov-
ered. Robot soccer [6] provides a perfect test environment both for unilateral colli-
sion avoidance (in case of opponent players) as for cooperative avoidance (in case of
team mates).
It is conjectured, but not yet rigorously proven, that in the case of mutual avoid-
ance (1) any “unequal-delay” solution leads to a higher mean delay compared to the
minimal mean delay obtained by the “equal delay” cooperative solution, and (2) route
deviation at full speed can always outperform any solution with speed reduction.
Optimal reactive control for mutual avoidance provides a powerful heuristic for a
multiple vehicle avoidance strategy. For all conflicting vehicles the mutual, optimal
deviation course can be calculated. These courses can be recombined in a conserva-
tive and consistent manner to get a global solution that approaches the pair wise
optimal avoidance as much as possible. Combining and merging the individual in-
terests of multiple traffic participants needs further research.
References
1. de Groot, R.M.: Dynamic traffic control of free-navigated AGV’s. Master thesis, Univer-
sity of Twente (1997).
2. Hopcraft, J., Schwarz, T., Sharir, M.: On the complexity of motion planning for multiple
independent objects; PSPACE-hardness of the warehouse man’s problem, Journal of Ro-
botic Research 3(4) (1984) 76-88
3. Vloemans, V.P.M.: Autonomous mobile robot collision avoidance through central plan-
ning. Master thesis, University of Twente (2003).
4. Bennewitz, M., Burgard, W., Thrun, S.: Finding and Optimizing Solvable Priority Schemes
for Decoupled Path Planning Techniques for Teams of Mobile Robots, Robotics and
Autonomous Systems, Vol. 41 (2002)
5. Schoute, A.L., Bouwens, P.: Deadlock free traffic control with geometrical critical sections.
Proceedings CSN94, CWI Amsterdam (1994) 260-270 (see
http://wwwhome.cs.utwente.nl/~schoute)
6. MI20 robot soccer website, http://hmi.cs.utwente.nl/robotsoccer

! "#$%'&)(*+#(**("
,.- +("*/ - 0 - 132#(*.465"7+#8
9:<;>=@?BAC;>DEGFIHJH>:<;KAML@:;AC=BDNPORQK?@:="SUTVAMWIXYC;KAMT

Z\[^]`_Ca>bdce]gfihjlknmPoprqKs^t^bdu
vxwzyR{)|#}{~}'^yR@
y
?@:!OM:QH>Fr^:!OPTH>?@FQOP=^HJ;>FIBLH>FOM=FIH>OG:HKACBWIFJ?A<>xQOM;J;>:OP=@D@:=BQ:x7:<
Hd::=AM=:RFH>F=@Y FIOPWOMYPFQAMWORD@:WOPTJOCQlAMWIW:DHJ;>FrA;JH>FIH>:GR=AJ:!F= BFOPWIOPYPFr
QAMWB;KACF=@"AM=@D¡AQOC;J;>:OP=@D@F=@Y¢AMLH>OP=@OP*OPL@"L@WIH>FAMYP:=^H"H>:¤£VE¦¥3§@¨F=¡AM=
L@=@:<RBWOM;>:D©:=^Fr;>OP=@:=^Hª ?@F3QOM;J;>:OP=@D@:=@Q:6QAM=«:T¬OM;>AMWFIX:D AMABF :QH>Fr^:
AC@FI=@Y6T®;>OPH>?@:BFOPWIOPYPFIQlAMW7H>: §^!H>OA*ACH>?@:6AH>FQAMWnO¯DB:W7OPT)H>?B:QOM;J;>:<
OP=@D@F=@YxL@WIH>FAMYP:=^HH>:°*E¦¥3§ª ?@:D@Fr;>:Q<H>FOP= §^H>O6EG¥3§F±H>?@:ACH>?
:6ACH>FIXlAH>FOP=6OPTH>?@:±BFOPWIOPYPFQAMWORD@:WAC=BD6H>?@:±;>:<^:<;>J:D@FI;>:QH>FIOP=xEG¥3§H>O §^6QAM=
:#F=^H>:;J@;>:<H>:D!AMH>?@:#D@:;>FrPACH>FIOP=!OPT~A3FIOMH>:Q1?@=@FQAMWORD@:WOPTnH>?@:#L@=BD@:<;>WIFI=@Y*EG¥3§ª
?@:*WOMYPFQAMW²A;JH3OPT
H>?B:*EG¥3§³O¯DB:W+FACJ:D«OP=«WOPYMFQAMW+´@:<;>F=@YPzH>:3OMT
D@FIHJ;>FIBLH>:DxWOMYPFQTµOM;EG¥3§ª OF*OM;JHKAM=^H=@OMH>FIOP=@¶RF=^H>:=^H>FOP=AC=BD6;>:e:QH>FIOP=7¶^±FWIW
:#JL@:QHOPT7H>?B:#D@:<^:WOMB:=^HzOPT)AJL@FIHKACBW:6AH>?B:6ACH>FIQlACWT¬OM;>AMWFIXlACH>FIOP=7ª OH>?@FI
:=@D:QAM=;>:JOM;JH²H>O:<FIH>F=@Y#OM;J·OP=WOPYMFQAMWR´@7:<;>F=@YP+T¬OC;
QOROM:<;KACH>F=@Y#;>OC7OCH>AM=@D
ACYP:=^H>)±?@FQK?D@:lAMWI)±FrH>?3H>?@:U=@OMH>FIOP=OPTAMQ:<AM=@DH>FI:UD@:<7:=@D@:=^H~WOPYPFIQlACW^T¬OM;>L@WAMª
?@FIA;>FJ:+F=±?BACH+FI+QAMWIW:DYP:=B:<;>FQ\ORD@:WF=@Y@;>F=@QFIBW:\T¬OM;+AQOM=BHJ;>L@Q<H>FI^:\D@:JQ;>Fr@
H>FIOP=©OMT\JQ:=BAC;>FOP3OPT\OMBFIW:6AMLH>OP=@OPOPL@AMYP:=^H>x£¬;>OMOMH>1¨±±?B:<;>:*H>?@:WIORQlACWWOPYPFIQlACW
HKAH>:3ACQ:3OPT+AM=³AMYP:=^HF±ORD@:W:D"PA;>F®ACBWIF=D@::=@DB:=@Q:3OMT²H>?@:3QOM;J;>:7OM=BD@FI=BY
*Ol^::=^H>ª¸\OP=@Q:<;>=@F=@Y¹¯L@:H>FIOP=@zOPT²ACL@H>OM6ACH>:D;>:AMJOP=@F=@YAM=@D"D@:DBL@Q<H>FOP=F=!H>?B:
WIOPYPFIQlAMWHKACH>:#ACQ:OPT)EG¥3§6AM=@DQO¯OM:;KACH>FI=@Y;>OMOMH>\:QAM=AC@BWI!*:H>?@ORD@T®;>OP
ROPWFIQ\QOPLHKACH>FIOP=AM=@DQOP*BLH>:;+AMWIYP:<B;KAzF=3H>?@:\T®;KAM:<OM;J·3OPTWOPYPFIQlACW^´@:;>FI=BYM
AC=BD 6AM=^RiMAMWILB:D WOPYPFIQPª+º#L;=@:»YP:=@:<;KAMWUO¯DB:W
T¬OM;EG¥3§¦FIAMJ:D OP=$QACH>:YPOM;J
H>?@:OM;J*AM=@D6YP:=@:;KACWBQlAH>:YPOC;>FQAMW@J:6AM=^H>FQªº#TTµL;JH>?B:<;
AMJFIQ±FI=¯H>:<;>:H\FI=6EG¥3§FIH>?B:
FIJJL@:OMTUW:AC;>=@F=@Y±FrH>?¦;>::Q<HH>O"AC;JH>FI´BQF®AMW²=B:L;KAMW+=B:<HdOM;J·¦AC@BWFIQlAH>FOP=@H>O"AMYM:=^H
RH>:AC=BD¦;>OMOMH>FQª7¼ :*LBJ:A"=@:<»YM:OP*:HJ;>FIQP¶7QlAH>:YPOC;>FQAMWORD@:WF=@YAC@@;>OÂMQK?
±?@:<;>:W:AC;>=@F=@YQlAC=6:#TµOM;>6AMWIWIF=^H>:<;J@;>:H>:DxF=6H>:;>*\OPT7OM;JB?@FJªR¥#QQOM;>D@F=@Y3H>O
H>?@:QR:<;>=@:H>FIQB;>FI=@QFrWI:*OPTUT¬:AMJFrFIWFrHd^¶7FrHFFI=¦H>?B:Q:=¯H>:<;OPTUOPL;OC;J·³H>O"AMQ1?@FI:^:
FI*BW::=¯HKAH>FOP=@OPT)AMLH>OP=@OP*OPL@;>OMOMH>\AMJ:D!OP=xH>?@:B;>OC7OMJ:DxBFOPFI:<H>FQEG¥3§
*ORD@:WVª
¶ ¶ ¶ ¶ ¶ ¶ ¶ ¶ ¶ ¶ M ª

! " # $ %
¢{ >{n@{)|#{~
& ('
)*,+-*/. jp`feaKbdbeo021 *(3 oul]rcdfebda>fdfe]40(1 *(5 o1h6^jl[ *87:9 a*febd]<;¯ob fe]gfea*c hP[¯o ;^cea>=Vulpr]Io!]r[¦cdhM[¯o ;Pfe]<?
febeo[Âc @]rcecd]rjl[ *(BDC kjlbE6Z\[^]`_Ca>bdce]gfihF1)bdaKcdcG0 BDC kjlbE6H0I JJI *
I *LKM*DKOt^N [Cf 9 aKb *P hPsRaKbd[â>fe]Q? B [Cfejlprjluh o[R6 7 beo[^c4S t^[R?>fe]rjl[¯op B ;RaKbeofe]rjl[^c *UTWV4X Y<XAZ V4[]\ Y
^ X _a`bcedgfDhH\jilkm2bnigY<kporqnX Yskut vxwzy{fGiA\}|\}o~AYuYsk 0j;^t^s^p * ]r[mMaKpguk =¬jlbduCo[^]rq>]r[ûmChPc feaz@c )G I0
m ;¯ob fo
[ ²jMj CGc 0oc 9 ]g[ûlfejlH[ 0R * Pa* 0R;R; *)G = I0 )l I *
* */3 ]gfdfea>beoltâ>b *P prj ?]lza 6;a>Eb ?Kza ;Pfe]rjl[GcdhMcdfeza @c *X bf>|\}YWXO~A|ncedgYuY V<Zjdg|Rcrn}]cd>_M 0 )) 0
\j * R I } I 0I JJ )*
R* *(3 ]gfdfeaKbeotâKb *7:9 a*prjlcecjlkUcdaKprsk =VsRjlt^R[ 6¯olbd]gaKc felj olEb 6^co![â>t^bdj @jlprza ?Kt^prolb#f 9 aKjlb h³jlk
Ac ? 9 ]rqKj ; 9 bdaK[^]ro *TWV4X R Acdg_M 0I I J } I )l 0I JJ *
* *}3 ]gfdfeaKbeotâKb * @#s¯opIoR[ ?Ka~jlkPupg]IoQp =V[âKt^bdjl[¯op]r[CfeaKbe o ?>fe]gj[z]r[±cdhM[¯o ;^cdaKc : o ;jlcecd]rs^pra @za ? 9 =
o[^]rAc @'jlk^f 9 a ;¯of 9 j ; 9 hPcd]rjlprjluh±jlk¯s^Q] ;jlpIob 6^]rcdjEb 6â>b *LdgbfEX g[gVdg|nc4Vu]c 0 ) J <)Gl IJ 0 I JJ R*
*rR* 1)k®olpgqKulbeolk *B [Go6ulaK[âKbeop+[^jfe]rjl[³jlkUo 9 t^prp * 8 [ bcX _ \ cdE¡m2fE\[zc4V4[]\ Y(¢WdA\ >X |nVu|Z}o
Rk mg\ Y £EZfE\¤\}|¡¦¥-kA§¨\ |Z©wdE¡ zk kDªdA« c-\ |¡O¬©X |XAZ fE\>`nj-Vu|©R _OiAX}YpV4[ ^ X}_a`bce\ c4VX}|Ro
`nf>V®|jZd>fxqd>f>Y<\zZ§FVdg|RoLdg¯±°X f]² 0 )GR*
)k®oprqKulbeok [ ulaKj 3a>febd] ol[ 3fej Rjlprjlul] 1opbda1ocejl[^]r[û]r[bdjlsRjlfe] Kc
*rR* 1 *{B @ Q? R6 ; <? Q? * |R|\ Y rXD¬©\ c®j³
dg_ \}cV4[zL\}|¡ f>cV ´2[gV\}Y~A|ncedgYuYpVQZjdg|[]d 0 )G I })Gµ 0 )G *
µ*rR* 1 )k®oprqKulbeok a >j[ >a Pfjk
prjlu] 1op ¯sRaKbd]r[ûlco[ ¯sRaKbda ¦pgjul] Kolp Kjl[febdjpgpraKbdc [
*©7:9 F? R? z; Q? 2¶ R6¶ z6 <? 2? *
m2fAXG[]d]dA¡ Vu|Z ^ (°2· ¸¹¹¹lo bZb]caºG³E»}¸ o¸¹¹¹Go(hV}A¼d Zjd>oT,dgY ZVub_Ok _Od>f>V[A\ |x~]|R]cVucbced
XDm2j}]V4[zEo ~m ^ X |GzdgfAd>|[>dam2fEXG[>d]dA¡}V®|jZAorqnX Y<ku½ºl¾w¸¹¹»]lo``Hkut v¾³Et ¿¾kÀ¥-kp¬©kp¥rbniAX Vu
w4Á:¡jk 0IJJJ *
*rR* 1)k®oprqKulbeok j âKpr]r[û >j[^[â >fe]rjl[^]rcdf"[â>f +jlb Mc 1ofeaKuljlbd] 1op zulaKj a>febd] «oc a Jfec
*Â3 j6 Ã? g? E ? <? 0 @ <? E; z?
fej ob ^c j 3j jb ] pga1obd[^]r[û [
Ä l E6 Å 9 @ @ E; 9 Q? RÅ Æ *Ã m2fEXG[>d]dA¡}V®|jZ ^ (°2· ¸¹¹ ¾}o bZ bAcO»»}³E»tlo
¸¹¹}¾ oÀh{V}dE¼ Zd]oT,dgY ZVub_Ok _Odgf>V4[]\}|O~]|R]cVuc4bcedrXmj}]V4[zAo ~m ^ X |Gzd>f]d>|[>dm2fEXG[>d]dA¡ Vu|jZAo
qnX Y<kº}» vw¸¹¹gÇ}kÀ¥-kp¬©kp¥rbniAX Vu w4Á:¡jk 0IJJ R*
) J *rR* 1 )k®oprqKulbeok [ jul] Kolp 7]rsRaKbd]r[ûlc*o[ tPfej 3ofea a ^t >fe]gj[$]g[ lo [h µ_loprtâ
*ÈB ÊÉ <? aË R6 + @ z6Ì z6 R? 3 = g6
jlul] Kc*Z\cd]r[û b lj s^[âKb +oceaKc
É <? K N x *Í¢WdgÎlVu]c\¢WdA\ Y []\¡jd>_MV4\Ì¡d ^ Vdg|[zV4\ Ao¤dgf>Vd ¡d
¬©\ cedg_Ã Ï\ c4V4[]\ Ao/¢ ^ ¬ 0 µ 0I JJ R*
))*r3* "jMjlp ^bd] ûla [ [Cfebdj ^t >fe]rjl[fej t^pgfe]rolulaK[fUmChPc fea c j [ $]rpga>h Mm j[^c Bf
<6 <6 *{+ j6 R? 3 z@ *{ 9 ¦ ÑÐ :É A6H0
I JJjI *

!"#%$& '! ()$+*+,,&-/.
012+3 4"# 57682 $+*8 9: "#;$& <(
?= >A@CBED'FG/HJILKNM'O1PQI1H
RTSQUAVWYX[ZSQ\]X/^`_bac^]Z;Ued?X[SQW/f1gQhiSQ\egQSjAk%\ehml]SQWon[hmXqp2^`_bf?VrisQtedeWou
vxw<y{z}|#~z}~8?yE&y
BED%FD'IL@>{ED'K>1HIP>1>1&ILPTDILPDT@>1FED'PFO>1DF2FEDI]N@T@CBI1FEC
ED@ PD'4A¡4I2@'>1PPD'¢)>1FEFO2P>1>1ILBI£NFEO2BED@'>?F¤O?¥PQILN>1F
¢¦I1@'D;I1T
E>?I1NF
ILFBED§T>1P¢¦I1@'DI1%@>A>1I1NF)Ä©oF4>LBD P%T>1PE'ªABED§P>1>1TIL
I1O?FE
>D'£eD P¡«@'>?F¤O?¥PQILN>?F>?H%BD¬¢>1NFe¦ILPQI1D D'P-I>?Fe2>?H/BDT>1P¢¦I1@'D>?H/BD
P>L)>LxILP¨%®eQI1FE¦I]P¯EP>1KND'ILPNFOILPD
BEDEPD'@'FD'IL@P>L¦KD
ª&BD
Fe£eD'PD2{NFEDILN@'EP>1KND'+I1FE°BED2NFEO?¥EK±I]Po¡P>L¦KD
¨)²³@'K±I1@Ì1KcD BE>{!>
D'¢QILKNB´BEDµP>1>1IL¶N
BEDILEEP>eI1@QBÁ¡¸·D'F¦I'£¹»ºILPD'FeD'POE¨©YKNDI1E
>
I¼@'>?KND DKi¡½¢¡->?KN@«D'@ PEN>?F¸>?HBDµEPD'@'
NFED'IL@
>{EDK>?H-I1F¶ILP
ILF½H>1P4BD!¦I1H>LP4BED°PDÌ]DFe
>1H;BD°FA£?D'PD°NFED'IL@xEP>1KND'¨©oF
>LPD P>¾PD'EPD'DFe!ILF¶D'FePD°P>1>1!ILP¿@'K±I1
>?H¦I1@«NFeD'PD'¢>¾¤¦FE¶I
@'>?KD'D'K¡¸¢¡)>1KNN@@'KN>?D'´H>1PÀ>?KN¥N>1F¶>?H-BED«NFe£eD PDFD'IL@
EP>1KND'¨
Á#FEOI<oT>/¬¢>?FecP>1>1ILP
ªLÂ¨ ÂT¥@QBeD'PO?D PED'>?FE¢PQILD'-BED/EPF@'KD%>#I1@CBED'£eD
BEQI1¢½§B¸BEDBED'K¶>?HI¯@'>?¥D P°ILKNO?D EPQI¯¢¡¢D'ILEK¡FO¯BENµÃP]>1Ä FED P
¦I1D'D B>Ab¨cÅcI]D'PTD
ILD
>§FÆNFe£eD'¢NOeI]N>?FE'ªI1>?FEO«>1BED P'ª@>?FE¢P¥E@ NFEO
I>1PDx@'>?KND Ç«D¢2D ÇEI1KD?¨È©Y2¥PFED>?¥2BIL-BEN-QI1¢KDÌL->£eD P¡BILP
D PH>1PI1FE@DP>L¦KD4§BED'FBED°ED'O1PD'D°>?H;H±PDD'E>?>?HIP>L)>LI]PÉNFE@ PDÌLD'¨
²#FNFeD PD¢FEOx>1D P£?IL>?F!N<B¦I]§BEDD-FA£?D¢OeILN>1F§B>TD°IFIL¥PQI1KbI¡
>
@'>?FE¢P¥E@'<D'FE@CBEILP{§H>LP@'>?¥D'PI1KO?D PQI{¨
²Ê@'>?KD'D'K¡!EiË)D PD'FA#I¡>xPD'EPD'DFe#BD-FD'IL@>{EDKc>?H/IP>1>1;I]P8
>!KDÌ]PFBEDx@'>1PPD'¢>?FEFEO°P>L)>L2IL¾§iBBEDBD'K¼>1H<I¥EQILK¡@CBE>?D'F¾@'>?F¹
FED'@'>?FEN¢FD oT>1P!¥EFOI>`TD PH¥EKILP¤@'±I1KcFED¥PQI1KcFED oT>1PÆJ²#Ì#ÌTN2¥EK±I]>1P¨
BEILEEP>eI1@QB°ILKN>PDÌ1@QBED<>{>?F<KN§BD'FBED;FA¥E-D'P>?H¬¢>?Fe§NFE@ PDÌLD'¨
BED'D§D'Ç{D PND'FE@D'H>1PD'BD¦I1N/>?HBED§EDÌ>@>?¦FED§>1BILEEP>eI1@QBED§B
BED2BE>1D->x@'>1D§B!BED2FE@'PDI1NFEO4@'>?KD'Çio¡e¨©oFµIEKN>1IBEDI¢¥ED'Fe
Fx-¡O1P>?¥bª{Â#¨Í=?D PQILD'ªeT>LPeD>?FBEN%NEDI-¥ENFEO2D'KND'@'D'P>L)>LILPÊo¡{D%I1FE
BED@'>?¥EKN«ED>?FE¢PQILDB¦I]i;NT>1PBe§BENKD@'>?FENED'PFEO¥E@QBI4@'>?-NFIL>?FHÎ>1P
¥EDF@>1F@ PD DILEKNN@IL>?FE'¨
©oFµ>?¥PT>1PFO4O1P>?¥IxFED'³@'>?FEFD'@ N>?FEN¢N2¥EKNIL>1PxJÏKD'Ç®{ >A>?KQª§BEN@QBµ§NKK
EP>`£{NEDI#ÐD'ÇiKND>{>?KEHÎ>1P/2¥EK±IL>?F>?HI;KNILPO?D<@'K±I1>?HFD'¥PQI1KEFED'oT>1P2o¡{D'&BIL
BI£?DFED'¥EP>L¦Be¡>?KN>1O?N@I1K)P>{>1'ªEBI1D'DF
D £eD'KN>1D'I1FEN¦KDDFeD'¾T>LP
<F
EP>?OLPDCQ¨/²#F>`§FÑF>£eDK§I]BD'IL@ÌLK#I]EP>eIL@CBHÎ>1P4>{EDKNFEO¾FED oT>1P¯¢P¥E@ ¹
¥PD'BIL¥ED'D BE>{E#HÎP>?@ÌLD'O?>1P¡BED'>1P¡!I1FEµO?D'>?D P¡EP>`£{NEDIEPD@'ND
O1¥EDKNFED#HÎ>1P/BD§O?D'FED'P@N¦KDDFeQIL>?Fx>?H)FED'oT>1P{%NFBEDFED ¸2¥EK±IL>?F>{>?KJ¨
©Y@ÌLF°DB>§FµBED'>1PD'@ÌLKNK¡§B«I1FµILPO?¥ED'FA;H±P>?8@ILD'O?>1P¡BED'>1P¡B¦I]#BD
ILBEDI]N@I1KÈ>{ED'KÈBED'K>PD'E¥@'D-@>1KND Ço¡>?HKNDILPFENFEOxNF!@>?FE@ PD'D-2¥EK±IL¹
>?FE-§BBEN-¢¡¢D'¨ BDÐD'ÇiKNDx2¥EK±IL>?F¾>A>?K§NKKDTD'KNK/¥ED'¾HÎ>1P-BD
PDILD'Fe§>?HcFE¥E¢PNI1K}ILE¦KN@ILN>1FEP>1KND''¨

ª ª ª
ª !" ª # ª $
ª % ª & ª ')(* ª '" ª !) ¨
+¯ z ,zbEz}|#zÈ
-
.0/1.32 \eu`SQriSn .5476)89;:=<6)>?@9A<BDCE6=8;9FHGI. f1UeWohm\eu]SWQ3j J%DS KML}^ÒW N*PRQcSQWorihm\j -TS
U
U1.
V . Q . QdeIg W1t{SQWou`SQW .@X hi\4Vriu`^`WohmYX WeZ;hin[Ig WeSQ[ n ZWohmX[SQWohmdeZÊY_ de\ [W ]ehi_S ^`^`\ n[tAVOW N`SQhmXShi\eSQn<V`rmu]SteW[V`hmOn P
Ig WeSQb \ aTriShicg WLde\eu`n[nop1noX[SQZ;en d 2 \ 2 riu`^`WohmYX WeZ;higCVracWohmX[SQWohi^`\_N^]W<YX WeSf1^`rmlVtehirihÍX»p
^`_ 2 rmu]cS P
teW[VhigfLp?nYX[SQZnT^]_ Xf dAV'X[hi^`\ecn g .@2 S f dAV'X[hi^`\eSQn<ZV'YX WeSQZVX[hmgCV_S h*i)j?j -TSk
l jEUeU . j k h
Pmj U j .
d X \eu`rihiOn W-X[W[V\en[rV'X[hi^`\2hio\ n1Q . QdeIg W1t{SQWou`SQWQ1j p .q hi1\ N1riSr W dÎHS ]en . gj a%TW ^`\ te\eSQsW Q&VnoSQnV1\ ] 2 U P
UerihigQVX[hm^]\enQTj t}Wo^1gQSQDS ]ehi\eu`n^`_LYX WeS uq\LX[SQWo\AV'X[hi^`\AVrLa^`\e_NSQWoSQ\egQwS v jjxL}SCVWon^`3_ a%TW ^]\ te\eSyW Q&Vn[STn vij
-TS
SU j zwuYfea%j 2 denYX[WohV? j ^^`1\ ]e^`| \ {VYX WeSZ-V'X[higCVrEf1^1gQhmSX»}p ^SQgX[deWo0S J%^X[Sf1SQWohiSQnQ*j ~È^`r .1V
- j
a&VZ#teWo!h ]eu`Sk%\ehml]SQWon[hÍX»p t}WoSn[nCj -"S
S
U jeUeU . j P h . gj -"Sk)l.
j . Q . QdeIg W1t{SQWou`SQW . aTTW ^`\ te\eSQWÈtAVnoSQTn n 2 \-Vriu`^`WohmYX WeZhmgZ;SYX We ^ ]hi\U{^`rmp?\e^`Z;hVrA!h ]eSQV`reYX WeSQ^`WYp .
uq \ > 8;9)9AC}Bc<3GI9:)<6)>03=GI8RBDC_G1BY:=c@ ? :)GHBIeBY j&UAVu`SQn -"U h= V j V . R . z/SQ'h ]eSQr
t}deter . a^`Z;U . jeRT^ÒW ]eWoSQcg W]OX PRQc^`noX[^`\ Pm^V`\egQV`nYX[SQWQj -TS
U.
h . Q . QdeIg W1t{SQWou`SQW .x2 UeUerihigQVX[hm^]\en^` _ aTTW ^`\ te\eSQWtAVn[SQn/hm\\e^`\ Prihi\eSCVWgQ^`ZUed?X¢V'X[hi^`\AVru]S^]Z P
SX[WYp . 6)81BDCE6)89FI6)>o[Gm1BYFD8;Gr:0FD9;BD<389 xF@:c8`6)YBm1 4r 479;FIBcxBY* j - h1n )S Uk j -"S
Uk .
. Q . QcdeIg W1t{SQWou`SQWQyj a . ac^]rmrihi\enCj{V1\ ]Q . Zd?X[sQriSQW .s2 riu`SQteW[Vhig§ZSYX We*^ ]en%_N^]W/u`SQ^`ZSX[Wohig§WoSC=V P
n[^`\ehi\eu .xw<3<o4sBD¡)x¢:)C 8R£FH9; 3j jn U -
-TS j -TS
U
U.
¤.0/1. /3. aW[Vhiu .x¥Y<38Y:T FD8;9:)<8R:47:*¦Y:)89;FDGI.§2 ]1]ehino^]\ P q SQn[rmS| p t}deter . a^ . j -TS
U
¤.
k . z .X Ig N1Z;hiririSQW . a^`\egQSQWo\ehi\eu«YX WeSSZSQWou`hi\eu«Wo^`riS^]_u`SQ^`Z;SX[WYp¼hm\¯\eSQdeWo^`hi\e_N^]WoZV'X[higQn .
uqb \ ¨ 6)O6)>A>©Bc> ¨ Y:TFIBDGIGI9A<*ª«9A<¬@B O6)>`3=GI8BcC_G6)<¢:)C 8BccGYs®7FD¯=C_9A>A>!BDIs°r6)c8;CE6=<3<1
°r6 GY¯B}BY)GH ¨ O:TFcBYBY)9A< ª)GY£®5> GHBD¡T9;BD0F= ¨5 ¦c>©xr:)c8±;°@:)>A>'6)<I j -TSS
l.
U. z . t . t)V`der .14s:¦Y:)8 6)<39 >'6)8R:=cGI. {uR²¬t}WoSQn[nQjaVZ#teWo!h ]eu`x S {xVn[n[VÌg W1denoSXoX[nV1\ ]³^^`1\ ]e^`\j
-TS
U*V .
S.0/1. t}_±VrisQu`W[V_ .`´ \ZV'YX WeSQZ-V'X[higQV`rZ*^ ]eSQrihi\euhi\Wo^`t^X[higQn . uq\ wc89 `FH96)>¥Y<38RBD>A> 9!ª*BD<FcB³6)<
3=C}¦Y:)> 9F 6)81BcCE6)8;9FI6=>0¢:=C 8;9A< ª=01;¢6)> C}BD8[1 @;¢6)Co¦IBc>>®7)GH ¨ O:TFcBIBY)9A<*ª
GY
µ BYFD8 YB0r:)8BcG[9<¬¢:)C 8Bc@FH9;BD<FcBE¶"·1¶H
c9A< ª*BD|¸BDc>'6Tª j -TSS j .
-Tl.0/1. t}_±VrisQu`W[V_ .´ \2u`SQ^`ZSX[WohmgTV`1\ ]X[^]U{^`ri^ù`higCVrWoSCVn[^`\ehi\euhi\2Wo^`t{^`X[hmgQn .§w<1<6)> G0: 6)8*±
BDCE6=8;9FHG@6=<wc8;9 xFD9;6=>¥Y<38RBD>A> 9'ª*BD<FIB j -TS n V
k)S j -TU j -TS
Sk .
-
-
.0/1. t}_ÎVrisu]W[V_ . {*^ ]eSQrihi\eu;gQ^`\e\eSQgX[hi^`\ehinYX/\eSX K^ÒW N2noX[WodegX[deWoSQTn nAno^]Z;Su`SQ^`Z;SX[Wohig<V`1\ ]gCV'OX P
SQu`^`WohigCVr¦V`noUSgX[n .xw<3<6=> G@: 6)81BDCE6)89FHG@6)<Ew¥ j j ¤ n V
k)S j l- j V)l
l*V .
-"V .0/1. t}_±VrisQu`W[V_ . {*^ ]eSQrihi\euÆg^]\e\eSgX[hi^`\ehinoX\eSX K^ÒW N1Hn n<gCV'X[SQu`^`WohigCVrj<u`SQ^`Z;SX[Wohig«VnoUSQgX[n
d±X["^ K&VOW ]e
n v We^`Z^`Z;^]Wo1U WehmgrmSCVWo\ehi\e1u v g . uqM\ ¨ O:TFcBIBY=9<*ª
G¹¢y[ºyx» ¼½½)·=@ ª GY8}¾¾=±O¾¿"
¼½½=·) µ 9=OÀB ª BI[BD> ª
9 C}3wC}BDc9FI6=<}¥I<1GI8;9A8 8RB0: ¨ *=GI9FHGY¥ ¨ ¢:)<THBcIBc<FcB ¨ O:TFcBIBY)9A<*ª
GY
¸3:)>!;¶=¾)Á«m¼½½DÂ=Ã³ Ã ¦Y:)9AGE®7* j V
l
l h .
Controller Network for a Humanoid Robot
Edmund S hierer, Markus Wuerzl
Vienna University of Te hnology
Institute of Handling Devi es and Roboti s
Favoritenstrasse 9 { 11, 1040 Vienna
Tel.: +43 1 58801 31815, FAX: +43 1 58801 31899
email: s hiererihrt.tuwien.a .at, wuerzlihrt.tuwien.a .at
1 Introdu tion
Humanoid robots presently are one of the most ex iting, most hallenging and
most attra tive resear h targets in the eld of mobile roboti s. It overs not only
me hani al engineering, but also in ludes problems from ele tri al engineering,
omputer s ien e, arti ial intelligen e and more often psy hology.
Me hani al engineering solves questions of onstru tion, me hani al loads
and properties whilst ele tri al engineering gives answers to problems like drive
systems and omponents, power supply and interfa ing high power devi es to
pro essors and ontrollers. Computer s ien e on the other hand des ribes algo-
rithms for motion ontrol, feedba k ontrol systems, pattern re ognition, image
pro essing and so forth. Arti ial intelligen e as a spe ial appli ation of om-
puter s ien e may be used for high-grade pro essing of various sensory inputs
and environmental in uen es to perform a more sophisti ated and onvin ingly
behaviour.
One of the keyroles in su h a omplex system like humanoids is the pro essing
system for motion ontrol to guarentee smooth and "realisti " traje tories.
2 The Buildup of the Body
The basi design of a humanoid robot follows the well proven onstru tion of
the human body. That is to say it has 2 legs, a torso, 2 arms and a head.
There is a ertain limit for the number of DOFs a humanoid robot should have
to satisfy fundamental requirements on erning motion-sequen es. A robot like
Pino [1℄ has 26 DOFs, what is in my opinion the minimum for a realisti nature
of motion (Figure 1).
3 Ele tri al Drives, Gears and Sensors
Ea h DOF is realized by one motor. Therefore Pino for example has 26 motors
to generate adequate movements of the robot. To keep the overall osts for
building a humanoid as low as possible standard servo motors an be used.
They are available at omparatively low pri es and fulll most of the items in
Fig. 1. Te hni al view of Pino ( ourtesy of Kitano Symbioti Systems Proje t) [1℄
the atalog of requirements for a ele tri al drive used in a humanoid robot. One
weak point in su h a standard hobby servo is the gear box often made from
syntheti material whi h should be hanged to metal toothwheels to improve
life y le and torque.
Mobile robots, in parti ular a humanoid, needs a bun h of various sensors.
Omitting high-level sensors like visual sensors for obje t re ognition and image
pro essing there is need for
{ equilibrium sensors to balan e the robot in any arbitrary situation
{ for e sensors to dete t ollisions with obje ts and obsta les or onta t to
ground
{ angular positioning sensors for determing joint positions
and mu h more.
4 Master Pro essor and Controller Network
A ontrol system for a humanoid robot shall onsist of several "levels of im-
portan e". Like in a human being one level uses the features and pro essing
apabilities of its underlying level. If the superior level fails by any reason the
lower one an ontinue and performs its asso iated fun tionality. This kind of
intera ting levels of omplexity is known as subsumption ar hite ture and guar-
entees operability of the system down to a ertain limit even if parts of the
ontroller suer a loss. A Low-budget-version of su h a network an onsist of a
master ontroller, where all the high-grade fun tions like path planning, image
pro essing and so forth are performed, and a bun h of low- ost ontroller nodes.
These nodes an then pro ess the ommands given from the master. The bus the
nodes are inter onne ted with is the widespread USB [2℄ due to its simpli itiy
and its availability in modern ontroller hips.
5 Con lusion
This paper des ribes our approa h to implement a ontroller network for a hu-
manoid robot like Pino [1℄. It makes it possible to extend step-by-step the pro-
essing apabilities of a robot only by adding new ontrollers to the network.
Referen es
1. Open PINO Platform, http://www.symbio.jst.go.jp/PINO/index.html

2. USB 2.0 Spe i ation, http://www.usb.org/developers/do s/
Programming by Integration in Robotics
José Luis Fernández-Pérez, Antonio C. Domı́nguez-Brito, Daniel

Hernández-Sosa, and Jorge Cabrera-Gámez
Instituto Universitario de Sistemas Inteligentes y Aplicaciones Numéricas en

Ingenierı́a (IUSIANI), Universidad de Las Palmas de Gran Canaria, Spain
jfernandez@iusiani.ulpgc.es, adominguez@iusiani.ulpgc.es,
dhernandez@iusiani.ulpgc.es and jcabrera@dis.ulpgc.es
1 Introduction
Developing software for controlling robotic systems is costly due to the complex-
ity inherent in these systems. There is a need for tools that permit a reduction
in the programming effort, aiming at the generation of modular and robust ap-
plications, and promoting software reuse. The techniques which are of common
use today in other areas are not adequate to deal with the complexity associ-
ated with these systems [1]. This document presents CoolBOT [2], a component
oriented framework for programming robotic systems, based on a port automata
model [3] that fosters controllability and observability of software components.
CoolBOT [2] is a C++ component-oriented framework for programming
robotic systems that allows designing systems in terms of composition and inte-
gration of software components. Each software component [4] is an independent
execution unit which provides a given functionality, hidden behind an exter-
nal interface specifying clearly which data it needs and which data it produces.
Components, once defined and built, may be instantiated, integrated and used
as many times as needed in other systems.
In CoolBOT, components are modelled as Port Automata [3]. This concept
establishes a clear distinction between the internal functionality of an active en-
tity, an automaton, and its external interface, sets of input and output ports.
Fig. 1 displays the external view of a component where the component itself is
represented by a circle, input ports, ii , by the arrows oriented towards the circle,
and output ports, oi , by arrows oriented outwards. Fig. 2 depicts an example
of the internal view of a component, concretely the automaton that models it,
where circles are states of the automaton, and arrows, transitions between states.
Transitions are triggered by events, ei , caused either by incoming data through
an input port, or by an internal condition, or by a combination of both. Double
circles indicate automaton final states. CoolBOT components interact and inter
communicate each other by means of port connections established among their
input and output ports. Data are transmitted through port connections in dis-
crete units called port packets. Port packets are also classified by their type, and

This work has been supported by the research project PI2003/160 funded by the
Autonomous Government of Canary Islands (Gobierno de Canarias - Consejerı́a de
Educación, Cultura y Deportes), Spain.
usually each input and output port can only accept a specific set of port packet
types.
CoolBOT introduces two kinds of variables as facilities in order to support
the monitoring and control of components: observable variables, that represent
features of components that should be of interest from outside in terms of control,
or just for observability and monitoring purposes; and controllable variables,
which represent aspects of components which can be modified from outside, in
order to be able to control the internal behavior of a component. Additionally,
to guarantee external observation and control, CoolBOT components provide
by default two important ports: the control port and the monitoring port, both
depicted in Fig. 3. The monitoring port: which is a public output port by means
of which component observable variables are published; and the control port,
that is a public input port through which component controllable variables are
modified and updated. Fig. 4 illustrates graphically a typical execution control
loop for a component using these ports where there is another component as
external supervisor.
i1 o1 e 10
c m
e 12
s3 s7 o1
e3 i1
e5 e9
e 11 e 14
s1 e1
s2 e2
s4 e6
s6 e 13
s8 s9
e4 e7
om in ok
in
s5
Fig. 1: e8
Fig. 3: The control
Component port, c, and the
external view Fig. 2: Component internal view
monitoring port, m
Internally all components are modelled using the same default state automa-
ton, the default automaton, shown in Fig. 5, that contains all possible control
paths that a component may follow. In the figure, the transitions that rule the
automaton are labelled to indicate the event that triggers each one, some of
them correspond to internal events: ok, exception, attempt, last attempt and fin-
ish. The other ones indicate default controllable variable changes: ns r , ns re , ns s ,
ns d , np, and nex. Subscripts in nsi indicate which state has been commanded: r
(running state), re (ready state), s (suspended state), and d (dead state). Event
np happens when an external supervisor forces a priority change, and event nex
occurs when it provokes the occurrence of an exeption.
The default automaton is said to be “controllable” because it can be brought
externally in finite time by means of the control port to any of the controllable
states of the automaton, which are: ready, running, suspended and dead. The
rest of states are reachable only internally, and from them, a transition to one
of the controllable states can be forced externally. Having a look to Fig. 5 we
can see how CoolBOT components evolve along their execution time. Basically,
the default automaton organize the life of a component in several phases which
correspond to different states: starting, devised for initial resource allocation;
ready, the component is ready for a task execution; running, here the component
executes its specific task; suspended, execution has been suspended temporally;
end, a task execution has just been completed.
Furthermore, there are two pair of states conceived for handling faulty sit-
uations during execution, which are part of the support CoolBOT provides for
error and exception handling. One of them devised to face errors during resource
allocation (starting error recovery and starting error states), and the other one
dedicated to deal with errors during task execution (error recovery and running
error states).
ns re
ns d
suspended
nc np
external
ns r
supervisor
ns d ns s ns re
control monitoring nc
ok finish
starting ready ns r running end ns d dead
i1 o1 nc exception
component exception np nex
...
...
np
ns r
attempt
ok (∅ | timer)
om ns re
in np ok
starting
error running error
recovery attempt error last recovery
(∅ | timer) attempt
nc
Fig. 4: A typical last
attempt ns d
starting
component control loop error
ns d
Fig. 5: The Default Automaton.
2 A Simple Demonstrator
CoolBOT has been conceived to promote integrability, incremental design and
robustness of software developments in robotics. In this section, a simple demon-
strator will be outlined to illustrate how such principles manifest in systems built
using CoolBOT. The first level of this simple demonstrator is shown in Fig. 6
and it is made up of three components: the Pioneer which encapsulates the set
of sensors and effectors provided by an ActivMedia Robotics Pioneer robot; the
PF Fusion that is a potential field fuser; and a Joystick Navigation component
for managing the robot using a joystick. The integration shown in the figure
makes the robot to to avoid obstacles while going to a destination point. The
second and last level of our demostrator is depicted in Fig. 7. Note that the
systems adds two new components, the Sick Laser which controls a Sick laser
range finder and Scan Alignment that performs self-localization using a SLAM
(Simultaneous Localization And Mapping) algorithm [5][6].
Robustness is another important aspect in robotics. Focusing on this aspect
several tests were prepared to show the behavior of the system upon the malfunc-
tuion of any of its components. In the system of Fig. 7, the key components are:
the Pioneer, Sick Laser and the Scan Alignment components. Based on these
three components, two tests were made. The first of them showed the way the
system works whenever any of the components hangs, and the second one is re-
lated to the degradation of the system when the Sick Laser component stopped
running.
PF
FUSION PF
FUSION
sonar positions/
sonars/bumpers high/low priority sonar positions/
odometry commands sonars/bumpers
odometry reset high/low priority
odometry commands
odometry reset
PIONEER
odometry correction
PIONEER SICK
LASER
high/low priority
commands high/low priority
commands scan
odometry correction
JOYSTICK
NAVIGATION
JOYSTICK
NAVIGATION SCAN
ALIGNMENT
map
Fig. 6: The avoiding

level Fig. 7: The whole system
3 Conclusions
This document describes briefly CoolBOT, a component-oriented C++ program-
ming framework that favors a programming methodology for robotic systems
that fosters software integration, concurrency and parallelism, asynchronous ex-
ecution, asynchronous inter communication and data-flow-driven processing. The
framework also promotes a uniform approach for handling faulty situations.
References
1. Kortenkamp, D., Schultz, A.C.: Integrating robotics research. Autonomous Robots
6 (1999) 243–245
2. Domı́nguez-Brito, A.C.: CoolBOT: a Component-Oriented Programming Frame-
work for Robotics. PhD thesis, Dpto. Informática y Sistemas, Universidad de Las
Palmas de Gran Canaria (2003)
3. Steenstrup, M., Arbib, M.A., Manes, E.G.: Port automata and the algebra of con-
current processes. Journal of Computer and System Sciences 27 (1983) 29–50
4. Szyperski, C.: Component Software: Beyond Object-Oriented Programming.
Addison-Wesley (1999)
5. Lu, F., Milios, E.: Robot pose estimation in unknown environments by matching
2d range scans. Proc. IEEE Comp. Soc. Conf. on Computer Vision and Pattern
Recognition, Seattle, USA (1994)
6. Lu, F., Milios, E.: Globally consistent range scan alignment for environment map-
ping. Autonomous Robots 4 (1997) 333–349
A Homeostatic-Adaptive Approach for Controlling
Robotic Systems ?
Daniel Hernandez, Javier Lorenzo, Antonio Dominguez, and Modesto Castrillon
IUSIANI, ULPGC, Las Palmas G.C., Spain email: dhernandez@iusiani.ulpgc.es
Extended Abstract
In this work, we present a hybrid mechanism for controlling a mobile robotic sys-
tem which combines concepts from homeostatic control and adaptive behavior. The
homeostatic control is inspired by an emotional approach consisting of a set of artifi-
cial hormones [2] computed from pre-categorical sensory data and also from high-level
application results. The adaptive behavior is implemented by a fuzzy controller whose
rules dynamically modify several system parameters, using the same structure to con-
trol both low-level hormonal loops and high-level application tasks. The objective of
this proposal is twofold: guarantee an acceptable image quality keeping the perceptual
data into a homeostatic regime, and use the adaptive behavior to obtain a better resource
management and dynamic response.
Fig. 1. Homeostatic regulation mechanism Fig. 2. Set of camera parameters
In Figure 1 the outline of a homeostatic regulation mechanism is shown. The state of

the system is considered to fall into one of these categories: homeostatic, overwhelmed
and understimulated. The objective is to keep the system in the homeostatic regime
modifying its behavior toward this goal. In computer vision applications, for example,
where the environment changes from a controlled conditions E to E 0 (Fig. 2) the per-
formance of the system will be maximum for another set of camera parameters δ 0 . So if
the system does not own a mechanism to detect the new environment, its performance
will drop since it will continue using the initial parameter set δ, and we must rely on an
external agent to readjust the parameter set to δ 0 .
In our proposal, the homeostatic regulation is based on different hormones: some
directly obtained from the images (h luminance, h whitebalance) and others from the
?
This work has been partially supported by the Spanish Min. of Educ. and FEDER funds
(proj. TIN2004-07087), the Canary Islands Regional Government (proj. PI2003/165 and
PI2003/160) and the Univ. of Las Palmas (proj. UNI2003/10 and UNI2004/25).
results of the visual task (h size). The h luminance hormone, for example, is computed
from the luminance of the image by dividing the image into five regions similarly to the
method proposed by Lee et al. [3].These five regions allow us to get different AutoEx-
posure (AE) strategies according to the nature of the object of interest giving different
weights to the average luminance in each region. For computing the h whitebalance
hormone we adopt the Grey World [4] assumption which tries to make the average
amount of green, blue and red in the image equal, by adjusting the gain parameters.
At low level, control loops should be coordinated to take into account interdepen-
dencies, as several homeostatic loops executing simultaneously can produce side effects
on each other that make the settling times larger than if execution sequence is super-
vised. In other cases, simply it makes not sense executing some loops when others are
far out from their desired regime values (e.g. focusing on a very dark image).
On the other hand, active-vision [1] and mobile robotic applications are usually
implemented as a set of multiple periodic tasks executing concurrently on limited re-
sources systems. If not properly managed, this contention could lead to poor perfor-
mance, threatening system security or even blocking its operation, when competing by
CPU time, for example [5].
Fig. 3. Elements of the homeostatic adaptive regulation mechanism
As a consequence of this analysis, the basic homeostatic control described previ-

ously has been complemented with a higher level adaptive control in order to improve
performance and also to fit the special characteristics of mobile robotic systems. We
have designed a fuzzy adaptive control based on the configuration of each task in the
system as a periodic process, with a desired frequency of operation to be respected
whenever possible. A set of fuzzy rules take inputs from hormonal system to produce
commands that can modify execution parameters. To make the control effective two
types of actions are generated from fuzzy rules: quality and frequency commands. The
quality signals force the output generated by a task to have a certain quality level,
implicating normally a variation in the resource consumption. The frequency signals
command new operation periods to the tasks, allowing also a modification on resource
demands.The figure 3 shows the control scheme combining homeostasis and adaptation.
Several tests have been performed to evaluate the correct operation of the homeostatic-
adaptive mechanisms on a real mobile robotic application. The goal is to follow a line
traced on the floor under different lightning conditions making of two hormonal drives
regulating simultaneously luminance and white balancing hormones from the image.
At high-level, a control loop carries out the line following task processing the image
from the camera stabilized by the hormonal loops. A “stress hormone” is computed as a
function of the trajectory curvature. For the hormones the homeostatic regime is defined
in the range [-0.2,+0.2].
Three sets of fuzzy rules have been defined in the adaptive controller. A “relax-
ation rule set” reduces the frequency of the hormonal loops when they are inside its
homeostatic regime. An “image quality rule set” slows down robot motion and de-
grades high-level tasks when image hormones are far from their desired values. When
stress hormone increases, a “stress rule set” is used to guarantee robot security, as it is
approaching a curved zone.
Fig. 4. Example of homeostatic-adaptive control - Internal hormones
Figures 4 and 5 show some data extracted from a test performed on an circuit con-
sisting on two straight parallel segments connected by curved sections on both ends.
The collected data correspond to a complete lap around the circuit. Lightning varies
along the path, with a specially dark area near the beginning due to the existence of a
kind of tunnel that the robot must traverse, so that, without homeostasis the task fails.
In Figure 4, the luminance hormone values are represented together with execu-
tion frequency values for luminance and white balance loops. As it is shown, when
luminance hormone separates from homeostatic regime, luminance loop runs faster to
recover image quality as soon as possible, while white balance becomes slower. This
effect is specially relevant when traversing the dark zone, between seconds 75 and 100.
In homeostatic regime, white balance is allowed to execute at a higher frequency while
luminance loop gets relaxed, for example when the robot is on first straight segment
and first curve (seconds 100 to 125).
Figure 5 shows the stress hormone values, its effect on high-level task quality and
the translation on an execution command affecting translational velocity. As we can
see, the curved sections of the circuit provoke an increase in stress level that reflects in
a reduction on robot velocity, for example when approaching and traversing first curve
between seconds 105 and 130. At the middle of dark zone the robot near stops, waiting
for the recovering of image quality by hormonal loops.
Fig. 5. Example of homeostatic-adaptive control - Application hormone
References
1. J. Y. Aloimonos. Introduction: Active vision revisited. In J. Y. Aloimonos, editor, Active
Perception. Lawrence Erlbaum Assoc. Pub., New Jersey, 1993.
2. Dolores Cañamero. Modeling motivations and emotions as a basis for intelligent behavior. In
Johnson Lewis, editor, Proceedings of the First Int. Symposium on Autonomous Agents, pages
148–155, New York, 1997. ACM Press.
3. June-Sok Lee, You-Young Jung, Byung-Soo Kim, and Ko Sung-Jea. An advanced video
camera system with robust AF,AE and AWB control. IEEE Transactions on Consumer Elec-
tronics, 47(3):694–699, August 2001.
4. Harsh Nanda and Ross Cutler. Practical calibrations for a real-time digital onmidirectional
camera. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR
2001), 2001.
5. David B. Stewart and Pradeep K. Khosla. Mechanisms for detecting and handling timing
errors. Communications of the ACM, 40(1):87–93, 1997.
A Mathematical Formalism for the Evaluation of
C-space for Redundant Robots
Roberto Therón, Vidal Moreno, Belén Curto, and Francisco. J. Blanco
Departamento de Informática y Automática

Facultad de Ciencias - Universidad de Salamanca
Plaza de la Merced s/n 37008. Salamanca (Spain)
theron@usal.es, vmoreno@abedul.usal.es
bcurto@abedul.usal.es, jblanco@abedul.usal.es
Extended abstract
In simple terms the collision detection problem can be stated as how can one
have a computer determine when two objects are about to collide, or have col-
lided? The first approach is to evaluate this obstacle avoidance in the robot
workspace; this involves the direct kinematic computation for every robot point,
since the kinematic information is required and the movement planning needs to
be done for each link. After the work of Lozano-Pérez [1] the collision detection
can be performed in a simpler way: the planning problem is solved within the
Configuration Space (C-space). The key point in this approach is the fact that
the robot representation is reduced to a single point in this space, independently
of the type of structure [2]. Lozano-Pérez named this planning task findpath [1],
where a path that connects two points without collisions is found in the C-space.
There is a previous task, findspace[1], that must be performed before any path
can be built: obstacles must be represented in C-space, i.e., C-obstacles must be
computed.
Although some manipulators will benefit from the fact that C-space evalua-
tion algorithms will only depend on the dimension of the considered C-space, for
instance, PUMA-like or Stanford robots, it can become an intractable problem
when considering high-dimensional redundant robots.
In this scenario, the advent of new methods that enable simple and faster
evaluations of the C-space is a challenge. In this sense, Kavraki, in [3], proposes
a method, based on the convolution of the obstacles and a mobile robot, to eval-
uate the discrete C-space. In this method the computation time is independent
of the complexity and shape of the robot and the obstacles for a given resolu-
tion of representation. Although it is doubtless a great advance on the way to
simpler C-space evaluations, only 2D mobile robots are considered, thus lacking
in generality.
While a lot of effort has been done dealing with mobile robots, only a few
works [4][5][6][7] are concerned with articulated manipulators. However, the later
kind of robots are a key element if one wishes to find a general method.
In [8] a more ambitious method is proposed by Curto for many types of
structures that include both mobile and articulated robots, where the obstacle
representation in the C-space is established as the convolution product of two
functions that represent the robot and the obstacles, respectively.
Although this is a big step, the method is not general enough, since some
structures —such as redundant robots— can not be considered. A further step
can be taken by an analysis of the proposed general method as applied to artic-
ulated robots, since changing the reference system for each link would provide
a new method that simplifies the computation, thus reducing both the memory
needs and the computational complexity [9].
This abstract presents a new general method for obstacle representation in
the configuration space for redundant robots. The method is based on the an-
alytical deconstruction of the configuration space, i.e., the separated evaluation
of the C-space portion contributed by the collisions of each link in the kine-
matic chain. The systematic application of a simple convolution of two functions
describing each link and the workspace, respectively, is applied. In order to do
that, the transformation of the workspace among reference systems located at
one point of each link is needed; in this step a sound method, well-known in
robotics, is used. As a main advantage, only the free-collision configurations of
the previous links in the chain are examined when calculating the portion of the
C-space for a given link.
Concisely, it is necessary to choose a proper coordinate system to work with,
considering the particular mechanical structure of the robot and relating it with
the particular degree of freedom (fig. 1).
In this way, a robot A [10] is viewed as a set of r rigid objects.The kinematics
of this chain, i.e., the movement restrictions imposed by the joint to each element,
Ai —the degrees of freedom, DOFs—, would determine some regions of the C-
space for these objects.
Taking into account the Superposition Principle, the calculation of CB for
a robot A, a kinematic chain of r links, is done through the union of all the
CBk related to each of the elements of the robot. The computation of every
C-obstacle region must be done through the evaluation of the associated CBk
functions.
CBk (q1k , · · · , qsk ), ∀k ∈ {1, . . . , r} (1)
with {q1k , · · · , qsk } ⊆ {q1 , · · · , qm }, where {q1 , · · · , qm } are the DOFs associated
to the robot A. That is, being B the function that describes the workspace,
for the k-th element only the subset of configuration variables associated to
it are considered, and, analogously to what is proposed in [8], each of the
CBk (q1k , · · · , qsk ) functions is evaluated as follows
Z
Ak (q1k , · · · , qsk , x1 , · · · , xn )B(x1 , · · · , xn )dx1 · · · dxn (2)
When solving the integral (2), Ak (q1k , · · · , qsk , x1 , · · · , xn ), the function that
represents the robot, is difficult to evaluate, due to its dependency on all of the
DOFs related to itself and to the previous links in the chain. Thus, the purpuse
Fig. 1. Reference systems used in the kinematic chain
of the Deconstruction method is to reduce this difficulty by choosing the proper

frames.
In order to do that, let’s consider the robot formed by the kinematic chain
of figure 1. As one can see, following the Denavit-Hartenberg method [11], a
frame is associated with each link, placing the origin at the end of the link; the
orientation of axes depends on the position and orientation of the link.
Following the D-H procedure, the Deconstruction method proposes to use
the frame determined by the previous link for the k-th element. Thus, for the
link 1 the frame FA0 —which coincides with the workspace frame, FW — is used;
similarly, for the k-th element the frame FAk−1 will be used (figure 1).
It is important to point out that after an homogeneous transformation, the
position and orientation of the link Ak , expressed related to the frame FAk−1 ,
will only depend of its associated degrees of freedom, that is, (q(u+1)k , · · · , qsk ).
After this step, only the following evaluation is needed for the link:
A0k (q(u+1)k , · · · , qsk , x01 , · · · , x0n ) (3)
Finally, (2), which is used to calculate the C-obstacle portion pertaining to

the element Ak , becomes
Z
A0k (q(u+1)k , · · · , qsk , x01 , · · · , x0n )B 0 (x01 , · · · , x0n )dx01 · · · dx0n (4)
Now, after the proper frame is chosen, as it can be seen in (4), it is possible
to study individually each one of the links. This fact makes possible a great
simplification in the calculation of CB, since only those configurations of the
previous links that do not produce collision are considered for the evaluation of
a given link (e.g., when the robot collides with an object at the first link, all
the different configurations for the rest of the links in the chain are part of the
C-obstacle CBk ).
This way, the evaluation of obstacles for each link can be done easily (and
independently) by a convolution product.
With the proposed method the portion of C-obstacle related to a particular
link, CBk , is obtained evaluating the inverse Fourier transform of the following
expression:
R h i
F Ā0 (0,···,0,q(v+1) ,···,qs ) (q(u+1)k ,···,qvk ,ξ(v+1)k ,···,ξn )
k k
(ξ1 ,···,ξv )
k
0
·F [B (ξ(v+1)k ,···,ξn )] dξ(v+1)k ···dξn
(ξ1 ,···,ξv )
k (5)
Subindices (ξ1 , · · · , ξvk ) denote vk -dimensional transform of functions defined
in Rn . Over the rest of coordinates, (ξ(v+1)k , · · · , ξn ), where a convolution can
not be found, the integration is performed.
References
1. Lozano-Pérez, T.: Spatial planning: A configuration space approach. IEEE Trans-
actions on Computers 32 (1983) 108–120
2. Udupa, S.: Collision Detection and avoidance in controlled manipulators. PhD
thesis, Department of Electrical Engineering, California Institute of Technology
(1977)
3. Kavraki, L.E.: Computation of configuration space obstacles using the fast fourier
transform. IEEE Tr. on Robotics and Automation 11 (1995) 408–413
4. Maciejewski, A.A., Fox, J.J.: Path planning and the topology of configuration
space. IEEE Tr. on Robotics and Automation 9 (1993)
5. Newman, W., Branicky, M.: Real-time configuration space transforms for obstacle
avoidance. The International Journal of Robotics Research 6 (1991)
6. Lozano-Pérez, T.: A simple motion-planning algorithm for general robot manipu-
lators. IEEE Journal of Robotics and Automation 3 (1987) 224–238
7. Lozano-Pérez, T., P.O’Donnell: Parallel robot motion planning. In: Proc. of the
IEEE Int. Conf. on Robotics and Automation. (1991) 1000–1007
8. Curto, B., Moreno, V., Blanco, F.J.: A general method for c-space evaluation
and its application to articulated robots. IEEE Transactions on Robotics and
Automation 18 (2002) 24–31
9. Therón, R., Moreno, V., Curto, B., Blanco, F.J.: Assessing methods for the eval-
uation of the configuration space for planar revolute manipulators. Journal of
Computational Methods in Science and Engineering 4 (2004) 149–156
10. Latombe, J.C.: Robot motion planning. Kluwer Academic Publishers, Boston, MA
(1991)
11. Denavit, J., Hartenberg, R.S.: A kinematic notation for lower-pair mechanisms on
matrices. Journal of Applied Mathematics (1955) 215–221
Global Modal Logics for Multiagent Systems:
A Logical Fibering Approach
Johann Edtmayr
University of Salzburg, Department of Computer Science,

Jakob Haringer Str. 2, A-5020 Salzburg, Austria
johann.edtmayr@cs.uni-salzburg.at
Extended Abstract
The problem of combining logics (or logical systems) is a growing area of interest
in modern logic as shown among other things in [4] and [3]. In recent years logic
is used more and more to formalize complex problems in artificial intelligence,
software engineering, and computer science as a whole. In particular the formal
modeling of systems of cooperating agents is of general interest, because in many
studies of complex systems, especially distributed systems, a multiagent model is
used. The formal modeling of the different aspects of systems in this area usually
requires combined systems of logics. A logical system modeling a reactive agent
should be a combined system of logics of knowledge, belief, time, and modal
logics (of action) [3]. In our logical modeling of a multiagent system (MAS) each
agent is represented by its own (local) logic. The global logical model of the
whole MAS is a composition of the local logics of the agents [6].
A very flexible approach for combining logics is the concept of logical fiberings
which was originally introduced by J. Pfalzgraf (cf. [4]). The development of this
approach was influenced on the one hand by the concept of polycontextural
logics, which comes from the philosophical work of Gotthard Günther, on the
other hand by the mathematical theory of fiber bundles. The concept of logical
fiberings provides a framework for the construction of complex logical models as
follows: Different logical systems called fibers, or local subsystems, are arranged
(modeled) over the points of a base space. Mathematically a logical fibering is
defined by a map π : E → B; E is called total space, and B the base space.
The fiber over a point b ∈ B is exactly the preimage set π −1 (b). Each fiber over
the base space can be a single logical system (e.g. a classical first order logic)
or itself a combined system of logics, i.e., a logical fibering. The base space of a
logical fibering can carry an additional structure, for example a communication
network, which represents some kind of interaction or communication between
the local subsystems. J. Pfalzgraf and coworkers have shown among other thing
in [7] and [6] that this flexible modeling language is also suitable to associate a
system of distributed logics to a MAS. In modeling a multiagent system (MAS)
in this way each individual agent obtains its own logic. In this context we speak of
the local logical state space (i.e. local fiber) of an agent. The whole logical model
can be formed by putting together all local fibers thus yielding the corresponding
logical fibering which presents the global logical state space of the MAS.
For the designer of an autonomous artificial agent the relationship between
knowledge and actions of the agent are of great interest. It has become customary
to use epistemic modal logic for the formal study of knowledge and beliefs of
such systems. So far the concept of logical fiberings was not used in conjunction
with modal logics and possible world semantics. A closer study of systems of
distributed modal logics in terms of fibered structures for the formal modeling
of multiagent systems seems to be reasonable (cf. [8]). Thus our research has been
aimed at testing whether it is possible to build a logical fibering, which represents
a global modal logic modeling a multiagent system, consisting of the local modal
logics assigned to the agents. It has turned out that such a logical fibering results
in a formal model of a MAS, which allows to study the knowledge and beliefs
of agents, taking into account the communication between the agents. To show
first results of our research we consider a simple task (manufacturing process)
which is performed by three cooperating robots. A global modal logic, in terms
of a logical fibering of local modal logics, is used for the formal modeling of the
multiagent system consisting of the three cooperating robots. The main result
of this work is that global modal logics can be represented as logical fiberings.
In our future work we intend to study distributed knowledge, common knowl-
edge, and awareness structures within logical fiberings of modal logics (cf. [2]).
A classical deduction problem can be treated symbolically as a corresponding
ideal membership problem using computer algebra support with the method of
Gröbner bases and the Stone isomorphism. This approach was already used for
m-valued logics parallelized by logical fiberings (cf. [5]). Based on existing work
([5], [1]) we intend to analyse if deduction problems in systems of fibered modal
logics can be treated in a similar way.
References
1. J. Chazarain, A. Riscos, J.A. Alonso, and E. Briales. Multi-valued logic and Gröbner
Bases with applications to modal logic. Journal of Symbolic Computation, 11:181–
191, 1991.
2. R. Fagin, J. Y. Halpern, Y Moses, and M. Y. Vardin. Reasoning About Knowledge.
MIT Press, 1995.
3. Dov M. Gabbay. Fibring Logics. Oxford University Press, 1999.
4. J. Pfalzgraf. Logical Fiberings and Polycontextural Systems. 1991. In: Proc. Funda-
mentals of Artificial Intelligence Research, Jorrand Ph., Kelemen J. (Eds.). Lecture
Notes in Computer Science 535 (subseries LNAI).
5. J. Pfalzgraf. On logical fiberings and automated deduction in many-valued logics
using Gröbner Bases. Revista Real Academia de Ciencias, Serie A de Matemáticas,
RACSAM (Royal Academy of Sciences of Spain) Special issue on Symbolic Compu-
tation in Logic and Artificial Intelligence, 98, 2004.
6. J. Pfalzgraf and J. Edtmayr. The concept of logical fiberings: Distributed logics for
multiagent systems. Cybernetics and Systems (EMCSR 04), 1:58–63, 2004.
7. J. Pfalzgraf, V. Sofronie, and K. Stokkermans. On semantics for cooperative agents
scenarios. Cybernetics and Systems (EMCSR 96), 1:201–206, 1996.
8. T. Porter. Geometric aspects of multiagent systems. Electronic Notes in Theoretical
Computer Science 81, 2003.
The H2 Control Problem: Comparison of State-space
and Transfer-function Approaches
Vladimír Kučera 1, 2, 3
1
Czech Technical University in Prague, Faculty of Electrical Engineering
2
Institute of information Theory and Automation, Academy of Sciences of the Czech
Republic
kucera@fel.cvut.cz
Extended Abstract
The H2 control problem consists of stabilizing the control system while minimizing
the H2 norm of its transfer function. Several solutions to this problem are available.
For systems in state space form, and under the standard regularity assumptions,
Doyle, et al. [1] obtained an optimal regulator in observer form by solving two
algebraic Riccati equations. Stoorvogel [6] relaxed the regularity assumptions while
Kučera [3] obtained a solution for systems in descriptor form. For systems described
by transfer functions, Park and Bongiorno [5] employed Wiener-Hopf optimization to
obtain the optimal regulator transfer function via operations with rational matrices.
Meinsma [4] and Kučera [2] derived a solution using operations with proper stable
rational matrices: inner-outer factorizations and proper stable projections.
The aim of this paper is to compare the two approaches. It is well understood that
the inner-outer factorization is equivalent to solving an algebraic Riccati equation.
However, why are the proper stable projections not needed in the state-space
approach?
To fix ideas, the H2 control problem in the “textbook” form is considered. Given a
state-space realization of the plant
⎡ A B1 B 2 ⎤
⎡G G ⎤
G = ⎢ 11 12 ⎥ := ⎢ C1 0 D12 ⎥
⎣G 21 G 22 ⎦ ⎢C D ⎥
⎢⎣ 2 21 0 ⎥⎦
find a proper rational controller R that stabilizes G and minimizes the H2 norm of the
transfer function H from w to z in the standard control system configuration
w z
G
u y
R
3 Supported by the Ministry of Education of the Czech Republic under Research Program
MSM212300013.
It is assumed that (A, B2) is stabilizable, (A, C2) is detectable, the matrix
⎡ A − jωI B 2 ⎤
⎢ C D12 ⎥⎦
⎣ 1
has full column rank for all finite ω, the matrix
⎡ A − jωI B1 ⎤
⎢ C D21 ⎥⎦
⎣ 2
has full row rank for all finite ω, and
T T
∆12 := D12 D12 > 0 , ∆21 := D 21 D 21 > 0.
First G is represented in terms of doubly coprime factorizations over RH∞
G = M −1 N = N M −1 ,
with the denominator factors block triangular
⎡ I M 12 ⎤ ⎡N N ⎤ ⎡N N ⎤ ⎡ I 0 ⎤
M =⎢ ⎥ , N = ⎢ 11 12 ⎥ , N = ⎢ 11 12 ⎥ , M = ⎢ ⎥.
⎣0 M 22 ⎦ ⎣ N 21 N 22 ⎦ ⎣ N 21 N 22 ⎦ ⎣ M 21 M 22 ⎦
Then all controllers RS that stabilize G are parametrized as
RS (W ) = ( X + WN 22 ) −1 (Y + WM 22 )
= (Y + M22 W )( X + N 22 W ) − 1
where X, Y and X , Y are RH∞ matrices that satisfy the Bézout equation
⎡ M 22 − N 22 ⎤ ⎡ X N 22 ⎤ ⎡ I 0⎤
=
⎢ −Y
⎣ X ⎥⎦ ⎢⎣ Y M 22 ⎥⎦ ⎢⎣0 I ⎥⎦
and where W is a parameter that ranges over RH∞.
Finally, two dual projection results are used to minimize the norm, in which the
asterisks denotes conjugation:
(1) Let E and F be matrices with equally many rows, with E in RH2 and F in RH∞.
Suppose that F is inner and F ∗E is in RH2⊥. Then for any RH2 matrix T,
2 2 2
E − FT = E + T .
(2) Let E and F be matrices with equally many columns, with E in RH2 and F in RH∞.
Suppose that F is co-inner and EF ∗ is in RH2⊥. Then for any RH2 matrix T,
2 2 2
E − TF = E + T .
Using doubly coprime factorizations,
H = G11 + G12 RS ( I − G22 RS )G21
= N11 − VN 21
where
V := M 12 ( X + N 22W ) − N12 (Y + M 22W ) .
~ ~
Write N 21 = UN 21 , where N 21 is co-inner and U is co-outer. Let P denote the
~∗
projection of N 11 N 21 on RH2. Then
~
H = H 1 − V1 N 21
~∗
with H 1 = N 11− PN 21 in RH2 and H 1 N 21 in RH2⊥. Therefore,
2 2 2
H = H1 + V1 .
Now
V1 = VU − P = N11K − N12W2
where
N11K := ( M 12 X − N12Y )U − P , W2 = WU .
~ ~
Write N 12 = N 12U , where N12 is inner and U is outer. Denote P the projection of
~∗
N 12 N 11K on RH2. Then
~
V1 = H 2 − N12V
~ ~∗ ⊥
where H 2 = N 11K − N 12 P is in RH2 and N 12 H 2 is in RH2 . Then
2 2 2
V1 = H2 + V .
To summarize,
2 2 2 2
H = H1 + H2 + V
provided V = U WU − P is strictly proper.

The optimal controller R0 corresponds to V = 0 , hence R0 = R S (U −1 P U −1 ) .
The state-space approach is based on expressing the doubly coprime factors of G
and R in terms of stabilizing state feedback and output injection gains. Let K and L be
matrices such that A+B2L and A+KC2 are stable; then
⎡ sE − ( A + KC 2 ) 0 K ⎤ ⎡ sE − ( A + KC 2 ) B1 + KD21 B2 + KD22 ⎤
M = ⎢ C1 I 0 , N=⎢
⎥ C1 D11 D12 ⎥
⎢ ⎥ ⎢ ⎥
⎢⎣ C2 0 I ⎥⎦ ⎢⎣ C2 D21 D22 ⎥⎦
⎡ sE − ( A + B2 L) B1 B2 ⎤ ⎡ sE − ( A + B2 L) B1 B2 ⎤
M =⎢ 0 I 0⎥, N = ⎢ C1 + D12 L D11 D12 ⎥
⎢ ⎥ ⎢ ⎥
⎢⎣ L 0 I ⎥⎦ ⎢⎣ C 2 + D22 L D21 D22 ⎥⎦
and
⎡ sE − ( A + KC 2 ) B2 + KD22 ⎤ ⎡ sE − ( A + KC 2 ) K ⎤
X =⎢ ⎥ , Y =⎢
⎣ −L I ⎦ ⎣ −L 0 ⎥⎦
⎡ sE − ( A + B2 L) K ⎤ ⎡ sE − ( A + B2 L) K ⎤
X =⎢ ⎥, Y =⎢ .
⎣ − (C 2 + D 22 L ) I ⎦ ⎣ −L 0 ⎥⎦
Take
T −1
K = −( B1 D21 + Q K C 2T ) ∆21
where QK is the largest non-negative definite solution of the Riccati equation
−1
AQ K + Q K A T + B1 B1T = ( B1 D21
T
+ Q K C 2T ) ∆21 T
( B1 D21 + Q K C 2T )T .
−1 −1
Then ∆212 N 21 is co-inner and N11( ∆212 N 21)∗ belongs to RH2⊥. Further take
−1 T T
L = − ∆12 ( B2 QL + D12 C1 )
where QL is the largest non-negative definite solution of the Riccati equation
−1 T
A T QL + Q L A + C1T C1 = ( B2T Q L + D12
T
C1 ) T ∆12 T
( B2 Q L + D12 C1 ).
−1 −1
Then N12 ∆122 is inner and ( N 12 ∆122 )∗ N 11 belongs to RH2⊥. Since N11K can be
−1
obtained from N11 by replacing B1 by K, ( N 12 ∆122 )∗ N 11 K belongs to RH2⊥ as well.
It follows that
1 2 1 1 2
2 2
H = N 11 + N 11K ∆21
2 + ∆12
2 W∆ 2
21
for any W in RH2. A minimum of the norm is achieved for W = 0, and

⎡ A + B2 L + KC 2 K ⎤
R0 = RS (0) = X −1Y = Y X −1 = ⎢ .
⎣ −L 0 ⎥⎦
The difference between the two approaches derives from a different construction
and use of doubly coprime factors. The transfer-function approach takes any fixed
doubly coprime factors of G, while the state-space approach parametrizes all such
factors in terms of K and L. This difference shows in full when the transfer function H
is manipulated so that the projection results can be applied. While the transfer-
function approach simply extracts the inner factor from N12 and the co-inner factor
from N21, the state-space approach shapes the two doubly coprime factors to make
them inner/co-inner by selecting K and L. This is achieved by solving appropriate
Riccati equations.
Now it is seen why no proper stable projection is needed in the state-space
approach. The process of shaping N12 and N21 results in trivial outer and co-outer
1 1 −1 −1
factors U = ∆122 and U = ∆21
2
. Consequently, the inner N12 ∆122 and the co-inner ∆212 N 21
−1 −1
cancel all stable dynamics when forming N 11 ( ∆212 N 21 )∗ and ( N 12 ∆122 )∗ N 11 K . That is
why P = 0 and P = 0 .
It is further noted that the design parameters K and L in the doubly coprime
factors of G directly define the optimal controller R0. Hence, the doubly coprime
factors need not be explicitly calculated.
References
1. Doyle, J.C., Glover, K., Khargonekar, P.P., Francis, B.A.: State space solutions to
standard H2 and H∞ control problems. IEEE Trans. Automatic Control 34 (1989)
831-847
2. Kučera, V.: The H2 control problem with internal stability. In: Proc. Joint CCA,
IASIC, and CACSD, Taipei (2004), to appear
3. Kučera, V.: The H2 control problem for descriptor systems. In: Proc. 16th IFAC
World Congress, Prague (2005), submitted
4. Meinsma, G.: On the standard H2 problem. In: Kučera, V., Šebek, M. (eds.):
Robust Control Design, A Proceedings Volume from the 3rd IFAC Symposium,
Prague, Czech Republic. Pergamon, Oxford (2000) 681-686
5. Park, K., Bongiorno, J.J.: A general theory for the Wiener-Hopf design of
multivariable control systems. IEEE Trans. Automatic Control 34 (1989) 619-626
6. Stoorvogel, A.A.: The singular H2 control problem. Automatica 28 (1992) 627-631
Analysis, Synthesis and Mimetization of Nonlinear
Dynamical Control Systems using Mathematica
Jesús Rodríguez-Millán, Carla González and Anna Patete
Universidad de Los Andes - Facultad de Ingeniería - Dept. Sistemas de Control

Apartado 11 - La Hechicera, Mérida 5101-A, VENEZUELA, Fax: 58-274-2402872
{jrmillan, carlag, patete}@ula.ve
Abstract.
In this work we report the development of computer-animated mimics of some
nonlinear dynamical control systems frequently used in the literature for
illustration purposes. Mathematically speaking these mimics are discrete-time
nonlinear dynamical control systems, arising from the discretization of
continuous-time nonlinear dynamical systems. Besides considering and
implementing standard linear discretization techniques, we also report some
experiments on nonstandard symbolic discretization of nonlinear dynamical
systems that exploit the advantages of both the Euler polygons method and the
Picard iteration method. The proposed nonstandard discretization techniques
might be described as a triplet (Taylor approximation, Euler sampling, Picard
interpolation), and therefore might be controlled at least threefoldly. In fact, we
go a little further by introducing controls over the sampling time and/or the
number of Picard iterations that depend on a prescribed fitting error criterion.
We used the integrated symbolic, graphic, and numeric capabilities of
Mathematica to develop all our work within a single computation environment.
Keywords. Nonlinear control systems, analysis, design, symbolic discretization,

mimetization, integrated computation, Mathematica.
1 Introduction
That the inner structure of our brain seemed to be better fitted to parallel image
processing than sequential numerical processing is summarized in the universal
saying a single picture worth a thousand words. We may roughly describe the content
of this work as a dynamic upgrade of the previous maxim: a mimic worth a thousand
pictures. More precisely, in this paper we report the development of a set of
computer-animated mimics of physical processes commonly used in control theory as
plants. These mimics allow visualizing both the open-loop dynamics of the plants,
and the performance, advantages, and disadvantages of control laws intended to
modify such open-loop behaviors. Comprehension through visual mimetization has
some obvious advantages over traditional numerical simulation, namely, it is closer to
physical and geometric intuition and, to a great extend, it does not depend on the
order of the systems, since what our brain recognizes is a unique dynamical moving
object. So, visualization allows us to exploit the strong geometric understanding and
integrating capabilities of our brains, instead of relying on their weaker algebraic
manipulation abilities. Contrastingly, visual mimetization might also jeopardize our
fine resolution capabilities concerning discriminating subtle details on the exhibited
dynamics.
2 Mimics as Discrete-Time Nonlinear Dynamical Systems
Having said that, it is worth to remark that our work is mainly concerned with the
dynamical aspects of controlled dynamical systems, and the mathematics supporting
it. Thus, beyond attractive graphics animation, for the sake of this work a mimic is an
nth-order discrete-time nonlinear dynamical control system (DT-NLCS) [1, 6],
Φ((k+1)T) = F(Φ(kT), u(kT)), Φ(0) = xo; Ψ(kT) = H((Φ(kT), u(kT)) (1)
obtained by discretization of a smooth enough nth-order continuous-time nonlinear
dynamical control system (CT-NLCS) [7]
x’ = f(x, u), x(0) = xo; y = h(x, u) (2)
Obviously, both the discrete-time state-vector orbit Φ and the output-vector orbit Ψ
depend on the discretization method. Besides the Euler, Euler-Picard and Euler-
Taylor-Picard discretization proposals already described in [6], in this paper we
consider modifying the Euler-Picard and Euler-Taylor-Picard discretization methods,
through the introduction of error-based controls over the sampling period and the
number of Picard iterations, to improve the computation time [2].
When the DT-NLCS (1) describes the closed-loop dynamics of a control system,
the trayectories Φ and Ψ also depend on the control laws u, which may have different
linear or nonlinear structures [3, 4, 5]. Anyhow, the associated controllers are linear or
nonlinear dynamical systems, and are themselves also discretized using the proposed
discretization methods.
At its present stage of development, our library of mimics includes a family of
pendulum system: a simple pendulum, an inverted pendulum, an inverted pendulum
mounted on a car, two single inverted pendulums mounted on a car, a double inverted
pendulum mounted on a car, a centrifugal pendulum, a one-dimensional airplane, and
a family of biophysical ionic-channels.
3 Standard vs Nonstandard Discretization Methods
In control engineering, approximated discretization of linear control systems [1] is

based on substituting first order derivatives by first order backwards differentials,
whereas dynamical systems people prefer forward differentials to approximate time
derivative to avoid invertibility problems in the nonlinear case. Yet, if thought of as
supports of approximate integration methods, both approaches lead to one-step Euler
polygons like approximate solutions for the CT-NLCS (2), as suggested in Figure 1.
In a translation to signal theory’s language, we could roughly think of Euler
polygons like integration methods as “periodic sampling plus linear interpolation”
methods, leading to periodic sampling plus linear interpolation discrete-time
trajectories, obtained by restricting Euler polygons to the sequence t = kT, for integers
k, where the length T of the one-step Euler polygon plays the role of the sampling-
period. We will refer to the so obtained discrete-time trajectories as the Euler
discretized trajectories, and to the associated discretization method as the Euler
discretization method.
For linear dynamical control systems, Euler discretization generates discrete-time
trajectories with good steady-state fitting properties, but not necessarily with good
transient-state fitting properties, provided the Shannon sampling theorem holds. In the
nonlinear case, however, it is usually necessary to use high multiples of Shannon
frequency of the associated linearized system to get similar fitting properties.
In the linear case we may improve transient-state fitting properties of Euler
discretized trajectories by increasing the sampling frequency. Yet, we might also
appeal to a different discretization method, namely, the exact discretization method
[1], where instead of discretizing the linear ordinary differential equation defining the
system, i.e., x’ = A x + B u, x(0) = xo, we apply zero-order hold to the control system
and sample the actual trajectory of the continuous-time linear system, to obtain the
discrete-time trajectory
x((k+1)T) = Ade x(kT) + Bde u(kT) (3)

where Ade = Exp(AT), and Bde = Integrate[0, T, Exp(As), ds]. Computing discrete-
time trajectories by exact discretization is computationally more expensive, but their
transient-state fitting properties are also considerably better. We could think of exact
discretization as “periodic sampling plus zero-order hold exact interpolation”.
In general, exact discretization is not extendable to nonlinear dynamical systems
because of the impossibility of computing their actual trajectories symbolically. Yet,
based on the Picard iteration method [0], we proposed the Euler-Picard discretization
method [6] which could be summarized as “Euler like periodic sampling plus Picard
interpolation”. The discretization strategy is also summarized in Figure 1, where it
can be compared to the Euler discretization strategy. It is well known that the Picard
iteration method holds locally, and that its fitting properties depends on the number of
iterations. So, for implementation purposes the number of computed Picard iterations
acts as a control mechanism, modifying both the computation time and fitting
properties of the discrete-time trajectories.
The Picard iteration method is an integrative method, hence usually prohibitively
expensive for CT-NLCS with complex nonlinearities, trigonometric ones for instance.
To simplify nonlinearities and improve computation times, we proposed in [6] to use
the Euler-Taylor-Picard discretization method where, instead of computing the Euler-
Picard discretized trajectories for the CT-NLCS (2), we calculate the Euler-Picard
discretized trajectories for the CT-NLCS x’ = F(k, x, u), x(0) = xo, where F(k, x, u)
denotes the kth-degree Taylor polynomial expansion of the f(x, u). This certainly
improves computation times; and moreover, in many cases where because of the
complexity the Euler-Picard discretization method fails, it succeeds on generating
good fitting discrete-time trajectories.
x
5
xe t
xt
4 xp t
t
To 2 To 3 To 4 To
Figure 1. Graphical comparison between standard Euler polygons discretization method vs the
proposed Euler-Picard discretization method. Curves x(t), xe(t) and xp(t) denote the solution of
the CT-NLCS (2), its Euler polygons approximation and its Euler-Picard approximation,
respectively. The discretized trajectories are obtained by restricting x(t), xe(t) and xp(t) to the
sequence of sampling times 0, To, 2 To, 3 To, … .
In this paper we go some steps ahead on improving the nonstandard discretization

methods previously proposed, by introducing two kinds of controls over them. On the
one hand, keeping the sampling period fixed we introduce a control over the number
of Picard iterations needed to satisfy a prescribed error criterion on each sampling
interval. This way we obtain a fixed sampling frequency varying Picard iteration
number discretization method. On the other hand, we fix an upper bound to the Picard
iterations number per sampling period, and introduce a control allowing to reduce the
sampling period when a prescribed error criterion is not satisfied after computing the
allowed number of Picard iterations. This experiments point towards non-fixed
sampling frequency discretization strategies for CT-NLCS.
The different non-standard discretization strategies proposed above are described
in detail and are applied to several case studies, and particularly to a family of
pendulum systems of interest in control and robotics. We study both typical
stabilization problems in control, like stabilizing mounted pendulums on the upper
vertical position, but also some oscillatory problems, to force the discretization
strategies to their limits.
Besides considering and implementing additional improvements on the proposed
non-standard nonlinear discretization methods, at the time being we continue working
on giving our proposals a sound mathematical foundation.
References
0. Coddington, E. and Levinson, N., Theory of Ordinary Differential Equations, Tata

McGraw-Hill Publishing Co., New Delhi 1977.
1. Isermann, R., Digital Control Systems, Volume 1, Fundamentals, Deterministic Control,
Springer-Verlag, Berlin, 1989.
2. Patete, A., Visualisación de Dinámicas No Lineales Mediante Mímicos en Mathematica,
M.Sc. Dissertation, Graduate Program on Control and Automation, Universidad de Los
Andes, Mérida, 2004.
3. Rodríguez-Millán, J., NLcontrol: a Symbolic Computation Toolbox for Nonlinear
Control by Extended Linearization, Proceedings of the 1998 American Control
Conference, Volume 3, 1390-1394, IEEE, Philadelphia, 1998.
4. Rodríguez-Millán, J., Analysis and Design of Nonlinear Dynamical Systems Using
Mathematica, Nonlinear Analysis, Theory, Methods and Applications, Volume 30,
Number 6, 3795-3805, 1997.
5. Rodríguez-Millán, J. and Cardillo, J., Symbolic Computing Aided Design of Nonlinear
PID Controllers by Extended Linearization, In F. Pichler, R. Moreno-Díaz and R.
Albrecht (Editors), Computer Aided Systems Theory, Lecture Notes in Computer
Sciences 1030, pp. 253-270, Springer-Verlag, Berlin, 1996.
6. Rodríguez-Millán, J. and González, C., Three Mathematica Supported Proposals for the
Discretization of Nonlinear Control Systems, to appear in the Proceedings of the 4th
WCNA-2004, Orlando, July, 2004.
7. Sastry, Sh., Nonlinear Systems, Analysis, Stability and Control, Springer-Verlag, New
York, 1999.
Embedded Control Systems: the Control Kernel
P. Albertos, M. Vallés and A. Crespo
Instituto de Autom¶atica e Inform¶ atica Industrial

Universidad Polit¶ecnica de Valencia
pedro@aii.upv.es, mvalles@isa.upv.es, alfons@aii.upv.es
Extended Abstract
The cost reduction in the hardware implementation of digital control systems
has two main efiects. On the one hand, the fleld of applications is enlarged as
the cost of the control system becomes afiordable with respect to the cost of the
controlled process. On the other hand, the control device is reduced on size lead-
ing to a limitation on the available resources. As a result, the control subsystem
is integrated into the whole controlled system. The control is embedded and the
main features of this kind of systems should be taken into account. The main
purpose of this paper is to review the characteristics of the embedded control
systems (ECS), as well as to outline some results the authors have developed
in difierent related issues. Among the main features, the following ones are the
most relevant from the control viewpoint:
1. Autonomy, to work independently of the workstation.

2. Fault-tolerance, to be able to work in a degraded or faulty mode.
3. Missing data operation, to cope with the absence of external information.
4. Reconflgurability, to be adapted to changing conditions. This requires a set
of control options as well as the transfer mechanism among these options.
5. Hybrid control, combining continuous time control and decision making sys-
tems.
6. Multitasking environment in the one-single-computer-based control.
7. Hard real time constraints in the control implementation.
8. Variable time delays in the communication as well as in the computation.
9. Resources optimisation (CPU, memory, energy, hardware redundancy, ...).
10. Safety, to avoid damages to people in the vicinity. Certiflcation is required.
These features require dealing with some issues in designing and implement-
ing the control, were many research groups are working on [12, 13, 11, 1, 16,
15, 22]. For these issues, some previous results of the authors are discussed and
referenced:
Non-conventional sampling and updating patterns . The signals sampling

and the control action delivering time instants should be appropriated for the
required controlled plant dynamics and the available resources. This implies
the option of difierent sampling rates (multirate control) and to consider the
lost of synchronicity. A review of this issue has been already discussed in [3].
Hybrid control systems . This control option is rather general and covers
many other applications. In our case, the discrete time component is related
to the decision about changes in the control structure and this will not
happen frequently. Then, the stability issue is not a real problem.
Sampling rate changes . It is well known that, in general, the quality of the
digital control decreases if the sampling period increases. Thus, if resources
are available, the control tasks period should be reduced and this implying
changes in the controller parameters as well as in the stored information.
The appropriated updating of controller parameters and internal data can
improve the transient behaviour of the controlled plant, as discussed in [6, 7].
Missing data control . Data availability is not at all guaranteed in any op-
erating condition. Thus, control algorithms should cope with missing data
situations. The common solution is to maintain the past values of the data,
but this solution is not appropriate if the data availability pattern is scarce.
The authors have been working in this issue and some solutions based on
virtual sensors implementations have been reported [20].
Event-triggered control . Most control algorithms are time driven. The sampling-
updating pattern is regular and synchronous, with a constant time interval
between actions (periodic actions). But dealing with harsh and uncertain
environments, some activities are triggered by external events. And these
events may happen randomly and at any time. A treatment based on the
missing data approach is possible and has been successfully implemented in
laboratory crane control [2].
Decision and supervisory control . Changes in the environment and avail-
ability of resources should be monitored and changes in the operation mode
should be decided by an upper level of control. There is a lot of literature
on the subject and most results can be easily adapted to be used in ECS.
Resources monitoring and control . Any autonomous system should adapt
its activities to the available resources. The monitoring and control of the on-
board batteries will result in forced changes in goals and control structures.
The real-time adaptive scheduling of the CPU will allow for the best possible
control at any operating conditions [9, 10, 17, 18].
Degraded and back-up (safe) control strategies . Changes in the operat-
ing conditions require changes in the controller but under any working con-
dition the system must keep some basic properties. In emergency cases, a
degraded behaviour should be accepted, ensuring the safety of the operation.
Reduced order models . Under normal operation, a detailed model of the
process and/or the controller can be used but it is well known that the
complexity of the controller is directly related to the complexity of the plant.
More properly, it is related to the complexity of the model used in the design.
Thus model reduction techniques should be considered to simplify either the
plant model or the designed controller, mainly to operate in degraded modes.
For that purpose, a control kernel should be implemented. This kernel, not
necessarily unique, should take care of the process under extreme working
conditions and will be fully developed.
And altogether, the integration of control design and computer control im-
plementation is crucial to achieve the best control performances [21]. To consider
the relevance of each control action, and consequently each control task, to mod-
ify accordingly their priorities and distribute the available computation time to
each one, to reduce the delivery time interval for any control action or to balance
the control performance degrading of all the control loops are features already
treated by the authors. Some results can be outlined from the references d com-
puter control implementation is crucial to achieve the best control performances
[4, 14, 5].
Control kernel
In real time embedded systems, an operating system kernel should be provided to
handle the basic requirements of the whole system even working in the emergency
conditions [8]. Similarly, a control solution should be implemented to ensure the
safe behaviour of the system even in the case of resources shortage. This implies
a reduced order model of the plan, a simplifled control algorithm and a reliable
operation. These basic properties should be captured by a kernel representation
of the system in order to apply an essential control. Thus, in this case, the sys-
tem will fall under this control until more resources (options) become available.
Among other emergency conditions, the following circumstances are analysed: i)
The time allocated for control computation is reduced and a safe operation (with
degraded performances) is foreseen. Thus, the sampling period may be enlarged
and only the slow part of the plant model could be considere, and ii) Under an
emergency condition, a quick reaction of the control is expected. The sampling
period is reduced and the slow behaviour of the plant may be considered as
unchanged.
In both cases a reduced model of the plant is desirable and a change in the
model parameters should be implemented. A similar reasoning will be applied if
dealing with the controller model. There are many model reduction techniques
but the real-time transfer between equivalent representations should be inves-
tigated. This is the main theoretical result of this paper. Generally speaking,
the kernel representation implies a reduced order model of the system and a
mechanism to transfer from a normal, extended model to the kernel and vice
versa. For that purpose, a simple algorithm to transfer between models, also
recovering the involved data, should be provided. The transferring approaches
require scheduling support to allow mode changes in the system. A mode change
is initiated whenever the systems detects a change in the environment or in the
internal state that must drive it from one operating mode to another allowing
the use of reduced models and the transfer between models. Several protocols for
mode changes can be found in the literature. A survey on mode change protocols
can be found in [19]. The main requirements to be achieved by the protocols are:
schedulability, periodicity, promptness and consistency.
References
1. Abeni, L., L. Palopoli, G. Buttazzo (2000) ”On Adaptive Control Techniques in

Real-Time Resource Allocation”.Proc. of the Euromicro Conf. on Real-Time Sys-
tems, Stockholm, Sweden. (2000).
2. Albertos, P., Sala, A. and Sanchis R. ”Digital Model Parameter Estimation with
Missing Data” IFAC World Congress. S Francisco. Vol N. pp 217- 222. (1996)
3. Albertos, P. and Crespo, ”A. Real-Time Control of Non-Uniformly Sampled Data
Systems”. Control Engineering Practice. Vol 7 (4). Pp: 517-522. (1999)
4. Albertos, P., A. Crespo, I. Ripoll, M. Valls, P. Balbastre ”RT Control Scheduling to
Reduce Control Performance Degrading.” 39th IEEE Conference on Decision and
Control CDC’2000 Sidney (Australia), (2000)
5. Albertos, P. and Crespo, A. ”Integrated Design and Implementation of Digital
Controllers. EUROCAST 2001. Lecture Notes on Computers Science. pp 385-392
Springer Verlag. (2001)
6. Albertos, P. , Valls, M.,and Valera, A. ”Controller Updating under Operational
Logic Changes”. European Control Conference. Cambridge (UK) (2003)
7. Albertos, P. , Valls, M.,and Valera, A. ”Controller Transfer Under Sampling Rate
Dynamic Changes” IFAC Workshop on Modelling and Analysis of Logic Controlled
Dynamic Systems. Irkutsk (Russia). (2003)
8. ARTIST. Advanced Real-Time Systems. Selected topics in Embedded Systems
Design. European Project IST-2001-34820. Y2 Reports. (http : //www.artist −
embedded.org/Roadmaps/ART ISTR oadmapsY 2.pdf ) (2004)
9. Balbastre, P., I. Ripoll, A. Crespo ”Control task delay reduction under static and
dynamic scheduling policies” In Proceedings of the 7th International Conference on
Real-Time Computing Systems and Applications, (2000)
10. Balbastre P., Ripoll I., Vidal J., Crespo A. ”A Task Model to Reduce Control
Delays”. Journal of Real-Time Systems. Volume 27, Issue 3, pp. 215-236.(2004)
11. Baruah S., Buttazzo G., Gorinsky S., Lipari G. ”Scheduling periodic task systems
to minimize output jitter”. Sixth Conference on Real-Time Computing Systems and
Applications. HK, December 13-15, Pp: 62-69. (1999)
12. Cervin, A. and Ecker, J. ”Feedback scheduling of control tasks”. IEEE Conference
on Decision and Control, CDC2000, Sydney, Vol. 1, Page 4871-4876. (2000)
13. Cervin, A. Henriksson, D., Lincoln, B., Eker, J. and Arzen, K-E. ”How Does Con-
trol Timing Afiect Performance?” IEEE Control System Magazine. (2003)
14. Crespo, A., I. Ripoll, P. Albertos ”Reducing delays in RT control: The control
action interval”. 14th IFAC World Congress. pp. 257-262, (1999).
15. Goebel R., Joo Hespanha, Andrew Teel, Chaohong Cai, Ricardo Sanfelice. ”Hybrid
Systems: Generalized Solutions and Robust Stability”. In Proc. of the 6th IFAC
Symp. on Nonlinear Contr. Systems. (2004)
16. Koutsoukos, X.D. P.J. Antsaklis, J.A. Stiver and M.D. Lemmon, ”Supervisory
Control of Hybrid Systems” (2000). Proceedings of IEEE,. (2000)
17. Masmano M., Ripoll I., Crespo A. ”TLSF: A New Dynamic Memory Allocator for
Real-Time Systems”. ECRTS 2004. Catania (Italy). (2004)
18. OCERA: Open Components for Real-Time Embedded Applications . IST 35102
European Project. European Commission. (http://www.ocera.org). (2002).
19. Real J., Crespo A. . ”Mode Change protocols for Real-Time Systems: A Survey
and a New Proposal”. Journal of Real-Time Systems. Vol 26 (2). pp. 161-197. (2004)
20. Sanchis, R. and P. Albertos . ”Recursive identification under scarce measurements.
convergence analysis”. Automatica 38, pp. 535-544. (2002)
21. Sanz, R. and rzn K-E.. ”Trends in Software and Control”.IEEE Control Systems
Magazine. June 2003. pp 12-15 (2003)
22. Seto, D., J. P. Lehoczky, L. Sha, K. G. Shin ”On task schedulability in real-time
control systems”. In Proceedings of the 17th IEEE Real- Time Systems Symposium,
pp. 13-21. Washington, DC., (1996)
Contribution to the Control of a Distributed
Solar Collector Field
Manuel Berenguel2 , Ryszard Klempous1 , Henryk Maciejewski1 , Maciej

Nikodem1 , Jan Nikodem1 , Imre Rudas4 , Loreto Valenzuela3
1
2
University of Almerı́a, Dept.Lenuajes y Computación, Ctra.Sacramento s/n,
04120 Almerı́a, Spain
3
CIEMAT, Platforma Solar de Almerı́a, Ctra.Senes s/n, E-04200, Almerı́a, Spain
4
Budapest Polytechnic, H-1034 Budapest, Doberdo u. 6, Hungary
Extended abstract
The purpose of this work is to analyze methods and algorithms to control the
ACUREX solar collectors field of the Plataforma Solar de Almerı́a - PSA). This
solar thermal system is based in parabolic trough technology. Parabolic trough
collectors are solar concentrating devices suitable to work in the 1500 C - 4000 C
temperature range. A parabolic trough collector is made up of a parabolic-
trough-shaped mirror that reflects and concentrates direct solar radiation onto
a receiver tube located in the focal line of the parabola. The concentrated radia-
tion heats a fluid flowing through the receiver tube, thus transforming the solar
radiation into thermal energy in the form of the sensible heat of the fluid. In the
ACUREX solar field the heat transfer fluid used is oil type synthetic which can
be heated up to 300 0 C. The heated oil is stored in an oil tank. This single tank
serves not only as feed tank, transferring oil from the power block to the solar
field, but also as thermal energy storage system, hot and cold oil can be stored
in the same tank because density of oil used change a lot with the temperature
and hot oil remains always in the upper levels of the tank because its density
is much lower than that of cold oil. Traditionally, as the energy source in this
system can not be manipulated, the way of controlling this system is by mean of
manipulation of the oil flow through the solar field in order to achieve a desired
oil temperature level at the outlet of the solar field.
The analysis of methods and algorithms to control the ACUREX solar field
is a starting point to formulate a model for solar collectors field control opti-
mization. The sole criterion of our optimization policy is to maximize the en-
ergy accumulated in the oil storage tank but the optimal strategy (maximum
efficiency of obtaining energy is a major factor) should fulfill some constraints
resulting from physical and technical restrictions. It results in a nonlinear dy-
namic optimization problem with nonlinear constraints, which is considered in
literature as hard.
References
1. ACUREX Field D.A.S. and Control System, Documentation version 2001,Platforma
Solar de Almerı́a, Almerı́a, Spain 2001.
2. Blanco-Muriel M., Alarcón-Padilla D.C., López-Moratalla T., Lara-Coira M.: Com-
puting the solar vector, Solar Energy 70(5):431-441, 2001.
3. Camacho E.F., Berenguel M., Rubio F.R.: Advanced control of solar plants, Ad-
vances in Industrial Control, Springer-Verlag, 1997.
4. Carmona R.: Analisis, modelado y control de un campo de colectores solares dis-
tribuidos con sistema de seguimiento en un eje, PhD thesis, Universidad de Sevilla,
1985.
5. Farkas I., Vajk I.: Experiments with internal model-based controller for Acurex
Field, IHP Programme, Proceedings of the 2-nd Users Workshop, Almerı́a, Spain,
2002.
6. Klempous R,. Maciejewski H., Nikodem M. , Nikodem J., Berenguel M., Valenzuela
. L.,” DATA DRIVEN METHODS AND DATA ANALYSIS OF A DISTRIBUTED
SOLAR COLLECTOR FIELD”, APLIMAT International Confernce Bratyslava,
February, 2005 (to be appear0
7. Solar thermal electricity generation, Lectures from the summer school at the P.S.A.
13-17 July 1998, Almerı́a, Spain.
8. Valenzuela L., Zarza E.: Example of module regulaci.c - software package for the
Acurex field. P.S.A. Doc.No. R06/97 LV, Almerı́a, Spain, 1997.
9. Valenzuela L., Balsa P.: A series and parallel feedfoward control schemes to regulate
the operation of a solar collector field, Serie Ponencias, Proceedings of the 1st Users
10. Vaz F., Oliviera R.: Neves de Silva R.: PID control of a solar plant with gain
interpolation. Serie Ponencias, Proceedings of the 1st Users Workshop, Almerı́a,
Spain, 1997.
11. Zarza E. at all: The DISS Project: Direct Steam Generation in Parabolic Trough
Systems. Operation and Maintenance Experience and Update on Project Status.
Journal of Solar Energy Enginnering Transactions of the ASME, 124(2):126- 133,
2002.
Author Index
Abalde, C. .................................. 71 Carrillo, V.................................. 80

Albertos, P. .............................. 321 Castrillón-Santana, M...... 215, 301
Alemán-Flores, M.................... 238 Ceska, M. ................................. 161
Alemán-Flores, P..................... 238 Chávez-González, A.M............. 143
Alonso, J. ......................... 134, 249 Choueka, Y. ................................ 60
Alonso, J.A......................... 83, 280 Collado, J.M. ........................... 265
Alonso, J.B....................... 197, 201 Combarro, E.F......................... 100
Alonso, M.A. .............................. 96 Conde, C. ................................. 259
Alonso-Ruiz, J.......................... 249 Crespo, A. ................................ 321
Álvarez-Leon, L. ...................... 238 Crespo, J. L................................ 58
Aparecida-dos-Santos, L.............. 3 Cull, P. ..................................... 235
Aquilino, C................................... 1 Curto, B.................................... 305
Aransay, J. ................................. 89 Da Silva, J.F. ............................. 92
Argüelles, L................................ 13 de Blasio, G................................ 17
Armingol, J.M. ......................... 265 De Lope, J. ............................... 255
Astola, J. .................................. 165 Del Aguila, I.M. ....................... 115
Augusto-Hoff, M. ......................... 5 De-la-Escalera, A. ................... 265
Bailón, A. ................................. 102 De-la-Mata, J..................... 87, 131
Ballarin, C. ................................ 89 Delgado, A.E............................ 261
Barcala, F. M............................. 55 Déniz-Suárez, O. ...................... 215
Barreiro, A............................... 150 De-Pedro, T. .............. 13, 249, 257
Benac-Earle, C. ......................... 77 Dershowitz, N............................. 60
Berenguel, M. .................... 15, 325 Di Crescenzo, A. ...................... 231
Bergasa, L.M. .......................... 251 Díaz, A. .................................... 127
Bernardos, P.............................. 58 Díaz, I. ..................................... 100
Beust, P.................................... 123 Díaz, V.J..................................... 80
Biernat, J. ................................ 178 Díaz-Prado, J.A.. ....................... 64
Blanco, A. .................................. 49 Domínguez, E............................. 55
Blanco, F.J............................... 305 Domínguez-Brito, A.C.212, 297, 301
Blanco, R. ................................ 150 Dubois, D.M............................. 155
Blauth-Menezes, P. ...... 3, 5, 23, 25 Edtmayr, J................................ 309
Born-Vieira, R. .......................... 23 Egiazarian, K. .......................... 165
Borrego-Díaz, J. ...................... 143 Enríquez, F................................. 80
Bosch, A........................... 102, 157 Esteban-Sánchez, M. B ............ 238
Breis, M. .................................. 194 Falcón, N. ................................ 201
Bubnicki, Z................................. 21 Fernández, M.A........................ 261
Buonocore, A. .......................... 229 Fernández-Caballero, A. ......... 261
Caballero-Gil, P. ............. 182, 184 Fernández-Pérez, J. L. ............. 297
Cabello, E. ............................... 259 Ferrer, M.A. ..................... 197, 201
Cabrera-Gámez, J. .................. 297 Fredlund, L. ............................... 77
Camino, A. ............................... 197 Freire, E..................................... 49
Campos, M............................... 136 Freire, J. L. .......................... 49, 69
Canton, M. ............................... 189 Fritz, A. .................................... 194
Caputo, L. ................................ 229 Fuentes-Pavón, R..................... 238
Cárdenes, R. ............................ 209 Fúster-Sabater, A..................... 184
Galan, M.................................. 219 Klancar, G. .............................. 284
Gamallo, P................................. 92 Klempous, R. ................ 15, 45, 325
García, A. ........................ 127, 269 Kopacek, P. .............................. 272
García, M. A. ........................... 269 Krena, B. .................................. 161
García, R. .................. 13, 249, 257 Kucera, V. ................................ 311
García-Barriocanal, E............. 139 Küng, J..................................... 113
García-Cabrera, L................... 119 Lee, Jeong-Eom ....................... 223
García-del-Amo, I.J. ................ 106 Lee, Seok-Joo ........................... 223
García-Lopez, F....................... 106 Licznerski, B............................... 11
García-Torres, M..................... 106 Lopes, G.P.................................. 92
Gervás, P. ........................ 127, 145 Lopez, J. ................................... 194
Giorno, V. ........................ 227, 233 López, M.T. .............................. 261
Girardi-Roggia, K. ...................... 5 López-Brugos, J. A................... 180
Gollmann, D. ........................... 186 López-Valles, J.M. ................... 261
González, C........ 13, 249, 257, 316 Lorenzo, J................................. 301
González, C........ 13, 249, 257, 316 Maciejewski, H............. 15, 45, 325
González-Cobas, J. D. ............. 180 Maravall, D.............................. 255
Gotchev, A. .............................. 165 Marín, D. ................................. 110
Graña, J..................................... 73 Marín, R. .................. 102, 136, 157
Grossmann, W. .......................... 41 Mariño, C................................. 205
Guerra-Artal, C. ...................... 215 Martín, J. ............. 43, 83, 194, 280
Guil, F...................... 102, 115, 157 Martín, J. A. ............................... 13
Guindos, F. .............................. 189 Martínez, J. .............................. 194
Gulías, V. ............................. 69, 71 Martín-Gorostiza, E................. 269
Halbich, C.................................. 29 Martín-Mateos, F.J. ........... 83, 280
Hamann, T.K.............................. 36 Martinucci, B. .......................... 231
Hernández-Goya, C. ................ 182 Matko, D. ................................. 284
Hernández-Sosa, J.D.212, 297, 301 Mauerkirchner, M. ..................... 47
Hervás, R. ................................ 145 Medina, N................................. 119
Herzog, K................................. 194 Melián-Batista, B. .................... 106
Hidalgo, M.J. ............................. 83 Menasalvas, E.......................... 110
Hilario, C................................. 265 Mira, J.................................. 1, 261
Hildebrand, L........................... 275 Miró-Julià, M............................... 7
Hoefer, G. .................................. 47 Mitterauer, B............................ 290
Holmberg, S. ............................ 155 Molina, A. ................................ 189
Horak, B. ................................. 277 Molinero, M. .............................. 55
Hurtado, M.V........................... 119 Mones, R .................................. 100
Ildefonso, T. ............................... 51 Montañés, E. ............................ 100
Isern-González, J. ............ 212, 215 Montes-Diez, R......................... 259
Izquierdo, R. ............................ 134 Mora, E. ............................. 58, 110
Jang, Min-Soo.......................... 223 Moraes-Claudio, D. ..................... 3
Jochinger, D. ........................... 173 Morales, A................................ 136
Juárez, J.M .............................. 136 Moreno, V. ............................... 305
Karba, R. ................................. 284 Moreno-Díaz, A. ........................ 43
Karpovsky, M........................... 169 Moreno-Díaz, R. .................. 17, 33
Kim, Sang-Jun ......................... 223 Moreno-Díaz jr, R...................... 33
Kim, Yong-Guk ........................ 223 Moreno-Pérez, J.A. .................. 106
Moreno-Vega, J.M. .................. 106 Rozenblit, J................................. 45
Moschner, M.............................. 41 Rubio, J. ..................................... 89
Munteanu, C. ........................... 219 Rubio Royo, E. ......................... 219
Naranjo, J.E............................. 257 Rudas, I. ................................... 325
Naranjo-Cabrera, A.M. ........... 212 Ruiz, J.L. .................................... 83
Nikodem, J. .................. 15, 45, 325 Ruiz-Alzola, J ........................... 209
Nikodem, M.......... 15, 45, 178, 325 Ruiz-Reina, J.L........................... 83
Nobile, A.G. ..................... 227, 233 Salort, J.................................... 136
Nuevo, J. .................................. 251 Sánchez, J. J............................... 69
Obitko, M. ................................ 277 Santana-Montesdeoca, J. M..... 238
Ocaña, M. ................................ 251 Santiago-Jorge, J. ...................... 69
Olivas, J.A. ........................ 87, 131 Scharinger, J...................... 45, 176
Orellana, F.J............................ 115 Schierer, E. .............................. 294
Otero, J. ..................................... 73 Schoute, A. L. ........................... 286
Palma, J........................... 136, 194 Schwaninger, M. ........................ 36
Palomo, M. .............................. 280 Segovia, J. ................................ 110
París, J....................................... 71 Serrano, A. ............................... 259
Park, Gwi-Tae ......................... 223 Serrano-Guerrero, J. ......... 87, 131
Patete, A. ................................. 316 Sicilia, M.A. ............................. 139
Penas, M .................................. 205 Sieniutycz, S. ............................ 245
Penedo, M.G ............................ 205 Smid, J...................................... 277
Pereira, G. ................................. 51 Snasel, V. ................................. 277
Pereira Machado, J. .................. 25 Sotelo, M. A...................... 251, 269
Pfalzgraf, J....................... 290, 292 Stankovic, R. ............................ 169
Pichler, F. .......................... 11, 173 Suarez Araujo, C.P. ................... 45
Piedra, J.A. .............................. 189 Szczurek, A. ................................ 11
Pirozzi, E. ................ 229, 231, 233 Szecówka, P. .............................. 11
Pitzer, E. .................................. 242 Tal, L.......................................... 60
Pogossova, E. .......................... 165 Therón, R. ................................ 305
Quesada-Arencibia, A................ 33 Toraño, J.................................... 13
Quintana-Domínguez, F. ......... 152 Travieso, C.M. ................. 197, 201
Ranilla, J.................................. 100 Troyano, J. A.............................. 80
Resconi, G................................ 134 Túnez, S.................................... 115
Reusch, B. ................................ 275 Valenzuela, L. .................... 15, 325
Ribadas, F.J............................... 96 Vallés, M. ................................. 321
Ricciardi, L.M.................. 227, 229 Vázquez, N ............................... 100
Ríos-Insua, D. ............................ 43 Vieira-Toscani, L. ........................ 3
Rivas, J. M. ................................ 13 Vilares, J. ................................... 96
Rivero-Martin, R...................... 106 Vilares, M................................... 73
Rodríguez, M.L. ....................... 119 Virto, M...................................... 43
Rodríguez-Aragón, L.J............. 259 Vojnar, T. ................................. 161
Rodríguez-Millán, J. ................ 316 Wagner, R. ............................... 113
Rodríguez-Rodríguez, A........... 152 Weiss, N. .................................. 275
Rodríguez-Rodríguez, J. C......... 33 Wuerzl, M................................. 294
Rosa, A..................................... 219 Zorrilla, M.E. ..................... 58, 110
Rossmann, J. ............................ 192
Roy, T....................................... 123
Instituto Universitario de Ciencias y Tecnologías Cibernéticas
Österreichische Gesellschaft Für Technologiepolitik

Extended Abstracts Book PDF

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Extended Abstracts Book PDF

Încărcat de

Drepturi de autor:

Formate disponibile

Alexis Quesada-Arencibia

Roberto Moreno-Díaz jr.

Cast and Tools for

10th Internacional Workshop on Computer Aided Systems Theory

It was agreed that EUROCAST international conferences would be organized every

The chairpersons, with the counselling of the International Advisory Committee,

I. Formal Approaches in Modelling

On the formal and physical frontier areas between human

Approximation Problems Categories.............................................................3

Computation of Partial Automata through Span Composition ......................5

Neural network sensitivity analysis applied for the estimation of

Fuzzy modeling for coal seams A case study for a hard-coal

Data Analysis of a Distributed Solar Collector Field ..................................15

Non-linear Formal Modelling of Amphibian Tectal Cells ..........................17

Optimization of a Class of Uncertain Systems Based on

Computational Simulation of Categorical Constructions ............................23

Composing Transitions Into Transactions in UML Diagrams.....................25

On the Application of Biological Vision Concepts and

II. Tools and Models for Social Systems

Theory-Building with System Dynamics: Principles and

Ontology Integration for Statistical Information .........................................41

Solving Influence Diagrams With Simulation Methods..............................43

Decision support system using data warehousing and on-line

Towards Automated Controlling of Human Projectworking

III. Intelligent Information Systems

On Recursive Functions and Well–founded Relations in the

Longest Sorted Sequence Algorithm for Parallel Text Alignment..............51

Information Retrieval and Large Text Structured Corpora .........................55

Matching with a Hierarchical Ontology ......................................................60

Business Intelligence Process through Information Extraction

Reasoning about a Functional Program for computing Gröbner

CheapTB: A Low Cost of Operation Distributed File System ....................71

Spelling Correction on Technical Documents.............................................73

Verification of Language Based Fault-Tolerance........................................77

Applying Stacking and Corpus Transformation to a Chunking

A Formally Verified Proof (in PVS) of tje Strong Completeness

Fuzzy ontologies generation in GUMSe .....................................................87

Extracting Computer Algebra programs from statements ...........................89

Cross-Language Classification of Function Words .....................................92

Integrating syntactic information by means of data fusion

An Iterative Method for Mining Frequent Temporal Patterns...................102

IV. Systems Applications Components

Data Mining with Scatter Search...............................................................106

Web Usage Mining Project for Improving Web-based Learning

Similarity Queries in Data Bases Using Metric Distances - From

A WEB-CASE Tool Prototype For Hybrid Software

An Augmentative Communication System based on Adaptive

A tool for thematic cartography of corpora ..............................................123

System-oriented Evaluation for Multi-tier Personalization of

The gaps of the thesaurus WordNet used in Information

Language for Uncertainty and Conflicting Agents....................................134

A Model-Based Architecture for Fuzzy Temporal Diagnosis ...................136

Ontology-Based Models of Human-Computer Interaction for

Extension of Ontologies assisted by Automated Reasoning

Applying Genetic Algorithms to Referring Expression

A Software Architecture for Effective Document Identifier

An Ontology for Reusing Synthetic Tasks ................................................152

Modelling and Simulation of Systems with Multi-level

A Tractable Subclass of Fuzzy Constraint Networks ................................157

Parallel State Space Generation and Exploration on Shared-

VI. Cryptography and Spectral Analysis

Tree-structured Legendre multi-wavelets..................................................165

A new Pseudo-Random Generator based on Gollmann Cascades

&' <